diff --git a/apps/fastvideo_studio/mock_server.py b/apps/fastvideo_studio/mock_server.py
index 2af805b91..98e241cf7 100644
--- a/apps/fastvideo_studio/mock_server.py
+++ b/apps/fastvideo_studio/mock_server.py
@@ -4,8 +4,8 @@ In-memory mock of the FastVideo Studio API (``server.py``).
 
 This mock implements the same ``/api`` routes and response shapes as the real
 FastAPI server but keeps everything in memory and never touches the FastVideo
-library, a GPU, or a database. It is meant to back the Playwright e2e suite so
-the Next.js frontend can be exercised end-to-end without a real backend.
+library or a GPU. It is meant to back the Playwright e2e suite so the Next.js
+frontend can be exercised end-to-end without a real backend.
 
 Job lifecycle is simulated by recording a start timestamp and *computing* the
 status on read: a started job reports ``running`` for a few seconds and then
@@ -13,14 +13,24 @@ flips to ``completed`` with an ``output_path``, so polling the job list / logs
 shows progression. Generated media is a tiny 1s ``testsrc`` MP4 built lazily
 with ffmpeg and cached.
 
+Pass ``--data-dir`` to seed jobs/datasets/settings from a real server's
+persisted SQLite database instead of the built-in fake demo data, so you can
+browse real job history through the mock without a GPU. This is a read-only
+snapshot taken once at startup: nothing created or changed while the mock
+runs is ever written back to that database, so it's safe to point at the
+same ``--data-dir`` a real ``server.py`` is using.
+
 Usage (from the ``apps/`` directory)::
 
     PYTHONPATH=.. python -m fastvideo_studio.mock_server --port 8190
+    PYTHONPATH=.. python -m fastvideo_studio.mock_server --port 8190 --data-dir ../outputs/ui_data
 """
 
 from __future__ import annotations
 
 import contextlib
+import json
+import logging
 import os
 import random
 import shutil
@@ -29,22 +39,31 @@ import tempfile
 import threading
 import time
 import uuid
+from pathlib import Path
 from typing import Annotated, Any
 
 from fastapi import FastAPI, File, HTTPException, UploadFile
 from fastapi.middleware.cors import CORSMiddleware
 from fastapi.responses import FileResponse, PlainTextResponse
 
-from fastvideo_studio.database import default_settings_dict
+from fastvideo_studio.database import Database, default_settings_dict
 from fastvideo_studio.models import (CreateDatasetRequest, CreateJobRequest, SettingsUpdate, UpdateCaptionRequest,
                                      model_label)
 
+logging.basicConfig(level=logging.INFO, format="%(asctime)s [%(levelname)s] %(name)s: %(message)s")
+logger = logging.getLogger("fastvideo.studio.mock")
+
 # --- Config -----------------------------------------------------------------
 
 # How long a started job stays "running" before it flips to "completed".
 COMPLETE_AFTER_SECONDS = 3.0
 FFMPEG_BIN = shutil.which(os.getenv("FASTVIDEO_FFMPEG_BIN", "ffmpeg"))
 
+# Marks a job dict as a frozen row loaded from a real --data-dir snapshot
+# (see _load_snapshot_from_db) rather than something this process is
+# simulating. Stripped from API responses by _public_job.
+_FROM_SNAPSHOT_KEY = "_from_db_snapshot"
+
 # A small catalogue of fake models keyed by workload type. Mirrors the real
 # server's {id, label} shape (label derived from the HF-style path).
 _MODELS_BY_WORKLOAD: dict[str, list[str]] = {
@@ -54,6 +73,10 @@ _MODELS_BY_WORKLOAD: dict[str, list[str]] = {
     ],
     "i2v": [
         "Wan-AI/Wan2.1-I2V-14B-480P-Diffusers",
+        # Real registered id (fastvideo/registry.py) -- needed so the frontend's
+        # `modelId.toLowerCase().includes('minimax-h3')` gate (speech-tag
+        # console, references, last-frame upload) actually lights up in the mock.
+        "MiniMaxAI/MiniMax-H3",
     ],
     "t2i": [
         "black-forest-labs/FLUX.1-schnell",
@@ -98,16 +121,76 @@ _mock_media_cache: dict[str, str] = {}
 _mock_media_lock = threading.Lock()
 
 
+def _testsrc_frame(width: int, height: int, index: int):
+    """A cheap, dependency-light stand-in for ffmpeg's `testsrc` pattern:
+    a gradient that shifts over `index` so an encoded clip visibly animates."""
+    import numpy as np
+
+    frame = np.zeros((height, width, 3), dtype=np.uint8)
+    frame[:, :, 0] = np.linspace(0, 255, width, dtype=np.uint8)[None, :]
+    frame[:, :, 1] = np.linspace(0, 255, height, dtype=np.uint8)[:, None]
+    frame[:, :, 2] = (index * 10) % 256
+    return frame
+
+
+def _build_mock_media_with_pyav(kind: str, path: str) -> None:
+    """ffmpeg isn't installed on every dev box; PyAV (a real FastVideo
+    dependency already, used for reference-video decoding) can encode the
+    same tiny clip directly via its bundled libav, no external binary."""
+    from PIL import Image
+
+    if kind == "png":
+        Image.fromarray(_testsrc_frame(320, 240, 0)).save(path)
+        return
+
+    import av as pyav
+
+    # h264 is what every browser's <video> element actually plays; mpeg4
+    # (MPEG-4 Part 2) is a PyAV-always-has-it fallback that no mainstream
+    # browser decodes, so gallery/job previews would just fail to load.
+    try:
+        pyav.codec.Codec("h264", "w")
+        codec_name = "h264"
+    except Exception:
+        logger.warning("PyAV has no h264 encoder available; falling back to mpeg4, which most "
+                        "browsers cannot play in <video>. Install ffmpeg for real preview playback.")
+        codec_name = "mpeg4"
+
+    container = pyav.open(path, mode="w")
+    try:
+        stream = container.add_stream(codec_name, rate=24)
+        stream.width = 320
+        stream.height = 240
+        stream.pix_fmt = "yuv420p"
+        for i in range(24):  # 1s at 24fps, matching the ffmpeg path's clip length
+            frame = pyav.VideoFrame.from_ndarray(_testsrc_frame(320, 240, i), format="rgb24")
+            for packet in stream.encode(frame):
+                container.mux(packet)
+        for packet in stream.encode():
+            container.mux(packet)
+    finally:
+        container.close()
+
+
 def _build_mock_media(kind: str) -> str:
-    """Build (once) a tiny testsrc clip of the given ``kind`` with ffmpeg; cache the path."""
+    """Build (once) a tiny testsrc-like clip of the given ``kind``; cache the path."""
     with _mock_media_lock:
         cached = _mock_media_cache.get(kind)
         if cached and os.path.isfile(cached):
             return cached
-        if not FFMPEG_BIN:
-            raise HTTPException(status_code=500, detail="ffmpeg is required to build mock media")
         fd, path = tempfile.mkstemp(prefix="fvstudio_mock_", suffix=f".{kind}")
         os.close(fd)
+        if not FFMPEG_BIN:
+            try:
+                _build_mock_media_with_pyav(kind, path)
+            except Exception as exc:
+                with contextlib.suppress(OSError):
+                    os.remove(path)
+                raise HTTPException(status_code=500, detail=f"PyAV fallback failed to build mock {kind}: {exc}") from exc
+            if not os.path.isfile(path) or os.path.getsize(path) == 0:
+                raise HTTPException(status_code=500, detail=f"PyAV fallback produced no mock {kind} bytes")
+            _mock_media_cache[kind] = path
+            return path
         if kind == "png":
             command = [
                 FFMPEG_BIN,
@@ -166,15 +249,23 @@ def _build_mock_media(kind: str) -> str:
 
 
 def _new_job_dict(req: CreateJobRequest) -> dict[str, Any]:
-    """Build a job dict (mirrors job_runner.Job.to_dict()) in the pending state."""
+    """Build a job dict (mirrors job_runner.Job.to_dict()) in the pending state.
+
+    Starts from req.model_dump() rather than re-listing CreateJobRequest's
+    fields by hand -- the previous version silently dropped name,
+    last_image_path, references, and every offload/tuning flag because it
+    was never updated as CreateJobRequest grew those fields. Spreading the
+    request means new fields flow through automatically.
+    """
     job_id = str(uuid.uuid4())
     return {
+        **req.model_dump(),
         "id": job_id,
-        "model_id": req.model_id,
-        "prompt": req.prompt,
         "workload_type": req.workload_type or "t2v",
         "job_type": req.job_type or "inference",
         "image_path": req.image_path or "",
+        "negative_prompt": req.negative_prompt or "",
+        "data_path": req.data_path or "",
         "status": "pending",
         "created_at": time.time(),
         "started_at": None,
@@ -182,17 +273,6 @@ def _new_job_dict(req: CreateJobRequest) -> dict[str, Any]:
         "error": None,
         "output_path": None,
         "log_file_path": None,
-        "num_inference_steps": req.num_inference_steps,
-        "num_frames": req.num_frames,
-        "height": req.height,
-        "width": req.width,
-        "guidance_scale": req.guidance_scale,
-        "guidance_rescale": req.guidance_rescale,
-        "fps": req.fps,
-        "seed": req.seed,
-        "negative_prompt": req.negative_prompt or "",
-        "num_gpus": req.num_gpus,
-        "data_path": req.data_path or "",
         "progress": 0.0,
         "progress_msg": "",
         "phase": "pending",
@@ -204,8 +284,13 @@ def _advance_job(job: dict[str, Any]) -> None:
 
     Status is *computed on read* from the recorded start timestamp, so polling
     the job list / logs naturally shows pending -> running -> completed.
+
+    Jobs loaded from a persisted DB snapshot (``_FROM_SNAPSHOT_KEY``) are a
+    frozen historical record, not something this process is simulating — skip
+    them so a real "pending" or "running" row doesn't silently start ticking
+    toward a fake "completed" just because someone polled it.
     """
-    if job["status"] != "running" or not job.get("started_at"):
+    if job.get(_FROM_SNAPSHOT_KEY) or job["status"] != "running" or not job.get("started_at"):
         return
     elapsed = time.time() - job["started_at"]
     if elapsed >= COMPLETE_AFTER_SECONDS:
@@ -219,9 +304,9 @@ def _advance_job(job: dict[str, Any]) -> None:
 
 
 def _public_job(job: dict[str, Any]) -> dict[str, Any]:
-    """Advance + return a copy safe to serialize."""
+    """Advance + return a copy safe to serialize (internal-only keys stripped)."""
     _advance_job(job)
-    return dict(job)
+    return {k: v for k, v in job.items() if k != _FROM_SNAPSHOT_KEY}
 
 
 _LOG_TAIL = [
@@ -346,6 +431,84 @@ def _seed() -> None:
     }
 
 
+# --- DB snapshot (--data-dir) ------------------------------------------------
+
+
+def _decode_references(value: Any) -> list[dict[str, Any]]:
+    """Reference lists round-trip through the DB as JSON text."""
+    if not value:
+        return []
+    if isinstance(value, list):
+        return list(value)
+    try:
+        decoded = json.loads(value)
+    except (TypeError, ValueError):
+        return []
+    return list(decoded) if isinstance(decoded, list) else []
+
+
+def _job_from_snapshot(row: dict[str, Any]) -> dict[str, Any]:
+    """Adapt a persisted job row into the mock's job dict shape.
+
+    Frozen as a read-only historical record (see _FROM_SNAPSHOT_KEY): the mock
+    never actually ran this job, so it doesn't get to simulate progress for
+    one. A row still marked "running" is orphaned -- the real JobRunner
+    reclassifies those as failed on its own restart, so mirror that here.
+    """
+    job = dict(row)
+    status = job["status"]
+    if status == "running":
+        status = "failed"
+        job["error"] = job.get("error") or "Server restarted (job was running)"
+        job["finished_at"] = job.get("finished_at") or time.time()
+    job["status"] = status
+    job["references"] = _decode_references(row.get("references"))
+    job["progress"] = 100.0 if status == "completed" else 0.0
+    job["progress_msg"] = "50/50 steps" if status == "completed" else ""
+    job["phase"] = "done" if status == "completed" else status
+    job[_FROM_SNAPSHOT_KEY] = True
+    return job
+
+
+def _load_snapshot_from_db(data_dir: Path) -> bool:
+    """Seed jobs/datasets/settings from a real server's persisted SQLite db.
+
+    Read-only: opens the db (Database.__init__ runs the same idempotent
+    schema migrations a real server startup would) and copies rows into the
+    mock's in-memory state; nothing the mock does afterward is written back.
+    Returns False (leaving the caller to fall back to _seed()) if no db file
+    is found at that path, rather than silently creating an empty one.
+    """
+    db_path = data_dir / "fastvideo_ui.db"
+    if not db_path.is_file():
+        logger.warning("--data-dir %s has no fastvideo_ui.db; falling back to fake demo data.", data_dir)
+        return False
+
+    db = Database(db_path)
+    settings = db.get_settings()
+    jobs = [_job_from_snapshot(row) for row in db.get_all_jobs()]
+    datasets = db.get_all_datasets()
+
+    with _state_lock:
+        _settings.update({k: v for k, v in settings.items() if v is not None})
+        _jobs.clear()
+        _jobs.update({j["id"]: j for j in jobs})
+        _datasets.clear()
+        _datasets.update({d["id"]: d for d in datasets})
+        _dataset_files.clear()
+        for d in datasets:
+            captions = db.get_dataset_captions(d["id"])
+            _dataset_files[d["id"]] = {"file_names": list(captions.keys()), "captions": captions}
+
+    logger.info(
+        "Loaded read-only snapshot from %s: %d job(s), %d dataset(s).",
+        db_path,
+        len(jobs),
+        len(datasets),
+    )
+    return True
+
+
 # --- App --------------------------------------------------------------------
 
 app = FastAPI(title="FastVideo Studio Mock API", version="0.1.0")
@@ -357,8 +520,6 @@ app.add_middleware(
     allow_headers=["*"],
 )
 
-_seed()
-
 
 @app.get("/api/__mock__")
 def mock_sentinel() -> dict[str, bool]:
@@ -423,6 +584,32 @@ async def upload_image(file: Annotated[UploadFile, File()]) -> dict[str, str]:
     return {"path": f"/mock/uploads/{uuid.uuid4().hex}_{os.path.basename(name)}"}
 
 
+ALLOWED_REFERENCE_VIDEO_EXTENSIONS = {".mp4", ".mov", ".mkv", ".webm", ".avi"}
+ALLOWED_REFERENCE_AUDIO_EXTENSIONS = {".wav", ".mp3", ".flac", ".m4a", ".ogg"}
+ALLOWED_REFERENCE_IMAGE_EXTENSIONS = {".png", ".jpg", ".jpeg", ".webp", ".bmp"}
+
+
+@app.post("/api/upload-media")
+async def upload_media(file: Annotated[UploadFile, File()]) -> dict[str, str]:
+    """Ref2VA reference upload (image/video/audio) -- the real server's
+    counterpart to upload_image, which only ever handles images."""
+    name = file.filename or "reference"
+    ext = os.path.splitext(name)[1].lower()
+    allowed = ALLOWED_REFERENCE_IMAGE_EXTENSIONS | ALLOWED_REFERENCE_VIDEO_EXTENSIONS | ALLOWED_REFERENCE_AUDIO_EXTENSIONS
+    if ext not in allowed:
+        raise HTTPException(status_code=400, detail=f"Invalid file type. Allowed: {', '.join(sorted(allowed))}")
+    if ext in ALLOWED_REFERENCE_VIDEO_EXTENSIONS:
+        media_type = "video"
+    elif ext in ALLOWED_REFERENCE_AUDIO_EXTENSIONS:
+        media_type = "audio"
+    else:
+        media_type = "image"
+    return {
+        "path": f"/mock/uploads/{uuid.uuid4().hex}_{os.path.basename(name)}",
+        "media_type": media_type,
+    }
+
+
 @app.post("/api/upload-raw-dataset")
 async def upload_raw_dataset(files: Annotated[list[UploadFile], File()]) -> dict[str, Any]:
     video_exts = {".mp4", ".webm", ".avi", ".mov", ".mkv"}
@@ -473,6 +660,9 @@ def create_job(req: CreateJobRequest) -> dict[str, Any]:
 
 
 def _start(job: dict[str, Any]) -> None:
+    # Once the mock is actually driving this job, it's no longer a frozen
+    # historical record -- let _advance_job simulate it like any other.
+    job.pop(_FROM_SNAPSHOT_KEY, None)
     job["status"] = "running"
     job["started_at"] = time.time()
     job["finished_at"] = None
@@ -552,6 +742,12 @@ def get_video(job_id: str) -> FileResponse:
         output_path = job.get("output_path")
         if job["status"] != "completed" or not output_path:
             raise HTTPException(status_code=404, detail="No output available for this job")
+    # A job loaded from a --data-dir snapshot may point at a real file this
+    # process never wrote, only reads -- serve it if it's actually there,
+    # otherwise fall back to the same synthetic clip every other job gets.
+    if os.path.isfile(output_path):
+        media_type = "video/mp4" if output_path.endswith(".mp4") else "image/png"
+        return FileResponse(output_path, media_type=media_type, filename=os.path.basename(output_path))
     # Mirror the real server: image workloads (t2i) output a .png served as an
     # image; everything else is a video.
     if output_path.endswith(".png"):
@@ -651,8 +847,18 @@ def main() -> None:
     parser.add_argument("--host", default="127.0.0.1", help="Bind address (default: 127.0.0.1)")
     # Default off the real server's 8189 so the mock never shadows a real API.
     parser.add_argument("--port", type=int, default=8190, help="Port number (default: 8190)")
+    parser.add_argument(
+        "--data-dir",
+        default=None,
+        help=("Read-only: seed jobs/datasets/settings from a real server's "
+              "fastvideo_ui.db under this directory (its --data-dir) instead "
+              "of fake demo data. Nothing the mock does is written back."),
+    )
     args = parser.parse_args()
 
+    if args.data_dir is None or not _load_snapshot_from_db(Path(args.data_dir).resolve()):
+        _seed()
+
     uvicorn.run(app, host=args.host, port=args.port, log_level="info")
 
 
diff --git a/apps/fastvideo_studio/src/app/scenes/page.test.tsx b/apps/fastvideo_studio/src/app/scenes/page.test.tsx
new file mode 100644
index 000000000..4474473e5
--- /dev/null
+++ b/apps/fastvideo_studio/src/app/scenes/page.test.tsx
@@ -0,0 +1,282 @@
+import { act, render, screen, waitFor, within } from '@testing-library/react';
+import userEvent from '@testing-library/user-event';
+import { afterEach, beforeEach, describe, expect, it, vi } from 'vitest';
+
+import ScenesPage from './page';
+import { getJobsList } from '@/lib/api';
+import { makeJob } from '@/test/factories';
+
+vi.mock('@/lib/api', () => ({
+  getJobsList: vi.fn(),
+  getJobVideoUrl: (id: string) => `http://test.local/api/jobs/${id}/video`,
+}));
+
+vi.mock('@/components/jobs/CreateJobModal', () => ({
+  default: (props: { readOnly?: boolean; editingJob?: { name?: string } | null; onClose: () => void }) => (
+    <div role="dialog" data-readonly={String(!!props.readOnly)}>
+      {props.editingJob?.name}
+      <button onClick={props.onClose}>close</button>
+    </div>
+  ),
+}));
+
+const wolf = (n: number, over: Parameters<typeof makeJob>[0] = {}) =>
+  makeJob({
+    id: `wolf-${n}`,
+    name: `wolf-lunch-clip-${String(n).padStart(2, '0')}`,
+    created_at: n,
+    num_frames: 240,
+    fps: 24,
+    prompt: [
+      'detailed_description:',
+      `[Shot 1] Wide shot. Clip ${n} opens.`,
+      `[Shot 2] At 00:02.000, the shot cuts to a close-up shot. (S1) <d>[English] Line from clip ${n}.</d>`,
+    ].join('\n'),
+    ...over,
+  });
+
+beforeEach(() => {
+  vi.mocked(getJobsList).mockReset();
+});
+
+describe('ScenesPage', () => {
+  it('groups clips into a scene and lists every shot in order across them', async () => {
+    vi.mocked(getJobsList).mockResolvedValue([wolf(2), wolf(3), wolf(1)]);
+    render(<ScenesPage />);
+
+    expect(await screen.findByRole('heading', { name: 'wolf-lunch' })).toBeInTheDocument();
+    expect(getJobsList).toHaveBeenCalledWith('inference');
+    expect(screen.getByText(/3 clips · 6 shots · about 0:30/)).toBeInTheDocument();
+
+    const rows = screen.getAllByRole('heading', { level: 3 }).map((h) => h.textContent);
+    expect(rows).toEqual(['wolf-lunch-clip-01', 'wolf-lunch-clip-02', 'wolf-lunch-clip-03']);
+    for (const n of [1, 2, 3]) {
+      expect(screen.getByText(new RegExp(`S1: “Line from clip ${n}\\.”`))).toBeInTheDocument();
+    }
+  });
+
+  it('numbers shots straight through the scene, not per clip', async () => {
+    vi.mocked(getJobsList).mockResolvedValue([wolf(1), wolf(2)]);
+    render(<ScenesPage />);
+    await screen.findByRole('heading', { name: 'wolf-lunch' });
+
+    const second = screen.getByLabelText('Shots in wolf-lunch-clip-02');
+    const badges = [...second.querySelectorAll('span.rounded-full.bg-secondary')].map((b) => b.textContent);
+    expect(badges.slice(0, 2)).toEqual(['3', '4']);
+  });
+
+  it('puts scene-wide times on each shot', async () => {
+    vi.mocked(getJobsList).mockResolvedValue([wolf(1), wolf(2)]);
+    render(<ScenesPage />);
+    await screen.findByRole('heading', { name: 'wolf-lunch' });
+
+    const second = screen.getByLabelText('Shots in wolf-lunch-clip-02');
+    // clip 2 starts at 10 s (clip 1 is 240 frames at 24 fps), and its second shot cuts in 2 s later
+    expect(within(second).getByText('00:10.000')).toBeInTheDocument();
+    expect(within(second).getByText('00:12.000')).toBeInTheDocument();
+  });
+
+  it('switches between scenes', async () => {
+    const user = userEvent.setup();
+    vi.mocked(getJobsList).mockResolvedValue([
+      wolf(1),
+      makeJob({ id: 'r1', name: 'rooftop-1', created_at: 99, prompt: '[Shot 1] Wide shot. A roof.' }),
+    ]);
+    render(<ScenesPage />);
+
+    expect(await screen.findByRole('heading', { name: 'rooftop' })).toBeInTheDocument();
+    await user.selectOptions(screen.getByLabelText('Scene'), 'wolf-lunch');
+    expect(await screen.findByRole('heading', { name: 'wolf-lunch' })).toBeInTheDocument();
+    expect(screen.queryByRole('heading', { name: 'rooftop' })).not.toBeInTheDocument();
+  });
+
+  it('hides the scene picker when there is only one scene', async () => {
+    vi.mocked(getJobsList).mockResolvedValue([wolf(1), wolf(2)]);
+    render(<ScenesPage />);
+    await screen.findByRole('heading', { name: 'wolf-lunch' });
+    expect(screen.queryByLabelText('Scene')).not.toBeInTheDocument();
+  });
+
+  it('shows status counts and marks a continuation clip', async () => {
+    vi.mocked(getJobsList).mockResolvedValue([
+      wolf(1, { status: 'completed' }),
+      wolf(2, { status: 'pending', prompt: '[Shot 1] Close-up shot. Continuing the same shot on <Subject 1>, still speaking.' }),
+    ]);
+    render(<ScenesPage />);
+    await screen.findByRole('heading', { name: 'wolf-lunch' });
+    expect(screen.getAllByText('1 completed').length).toBeGreaterThan(0);
+    expect(screen.getAllByText('1 pending').length).toBeGreaterThan(0);
+    expect(screen.getByText('continues previous shot')).toBeInTheDocument();
+  });
+
+  it('previews a completed clip on demand', async () => {
+    const user = userEvent.setup();
+    vi.mocked(getJobsList).mockResolvedValue([
+      wolf(1, { status: 'completed', output_path: '/out/a.mp4' }),
+      wolf(2, { status: 'pending' }),
+    ]);
+    render(<ScenesPage />);
+    await screen.findByRole('heading', { name: 'wolf-lunch' });
+
+    expect(screen.getAllByRole('button', { name: 'Preview' })).toHaveLength(1);
+    expect(screen.queryByLabelText(/Generated video/)).not.toBeInTheDocument();
+    await user.click(screen.getByRole('button', { name: 'Preview' }));
+    expect(screen.getByLabelText(/Generated video for wolf-lunch-clip-01/)).toHaveAttribute(
+      'src',
+      'http://test.local/api/jobs/wolf-1/video',
+    );
+    await user.click(screen.getByRole('button', { name: 'Hide video' }));
+    expect(screen.queryByLabelText(/Generated video/)).not.toBeInTheDocument();
+  });
+
+  it('opens a pending clip for editing and a finished one read-only', async () => {
+    const user = userEvent.setup();
+    vi.mocked(getJobsList).mockResolvedValue([wolf(1, { status: 'completed' }), wolf(2, { status: 'pending' })]);
+    render(<ScenesPage />);
+    await screen.findByRole('heading', { name: 'wolf-lunch' });
+
+    await user.click(screen.getByRole('button', { name: 'Edit' }));
+    let dialog = screen.getByRole('dialog');
+    expect(dialog).toHaveAttribute('data-readonly', 'false');
+    expect(dialog).toHaveTextContent('wolf-lunch-clip-02');
+    await user.click(within(dialog).getByRole('button', { name: 'close' }));
+
+    await user.click(screen.getByRole('button', { name: 'View' }));
+    dialog = screen.getByRole('dialog');
+    expect(dialog).toHaveAttribute('data-readonly', 'true');
+    expect(dialog).toHaveTextContent('wolf-lunch-clip-01');
+  });
+
+  it('timeline segments jump to their clip', async () => {
+    const user = userEvent.setup();
+    vi.mocked(getJobsList).mockResolvedValue([wolf(1), wolf(2)]);
+    render(<ScenesPage />);
+    await screen.findByRole('heading', { name: 'wolf-lunch' });
+
+    const scroll = vi.spyOn(Element.prototype, 'scrollIntoView');
+    await user.click(screen.getByRole('button', { name: /Go to clip 2: wolf-lunch-clip-02/ }));
+    expect(scroll).toHaveBeenCalled();
+    expect((scroll.mock.instances.at(-1) as unknown as Element).id).toBe('clip-wolf-2');
+  });
+
+  it('shows an empty state when there are no jobs', async () => {
+    vi.mocked(getJobsList).mockResolvedValue([]);
+    render(<ScenesPage />);
+    expect(await screen.findByText(/No inference jobs yet/)).toBeInTheDocument();
+  });
+
+  it('shows an error with a retry that reloads', async () => {
+    const user = userEvent.setup();
+    vi.mocked(getJobsList).mockRejectedValueOnce(new Error('API down')).mockResolvedValue([wolf(1)]);
+    render(<ScenesPage />);
+
+    expect(await screen.findByRole('alert')).toHaveTextContent('API down');
+    await user.click(screen.getByRole('button', { name: 'Try Again' }));
+    expect(await screen.findByRole('heading', { name: 'wolf-lunch' })).toBeInTheDocument();
+    await waitFor(() => expect(getJobsList).toHaveBeenCalledTimes(2));
+  });
+
+  it('says so when a clip has no shots', async () => {
+    vi.mocked(getJobsList).mockResolvedValue([makeJob({ id: 'e', name: 'empty-1', prompt: '' })]);
+    render(<ScenesPage />);
+    expect(await screen.findByText(/prompt has no shots/)).toBeInTheDocument();
+  });
+
+  describe('re-runs of the same clip', () => {
+    it('shows one clip with a take picker instead of a row per re-run', async () => {
+      vi.mocked(getJobsList).mockResolvedValue([
+        wolf(1, { id: 't1', created_at: 1, status: 'completed' }),
+        wolf(1, { id: 't2', created_at: 2, status: 'failed' }),
+        wolf(1, { id: 't3', created_at: 3, status: 'completed' }),
+        wolf(2, { id: 'n', created_at: 4, status: 'pending' }),
+      ]);
+      render(<ScenesPage />);
+      await screen.findByRole('heading', { name: 'wolf-lunch' });
+
+      expect(screen.getAllByRole('heading', { level: 3 })).toHaveLength(2);
+      expect(screen.getByText(/2 clips · 4 shots/)).toBeInTheDocument();
+      const picker = screen.getByLabelText('Take for clip 1') as HTMLSelectElement;
+      expect(picker.options).toHaveLength(3);
+      // newest completed take is shown by default (not the failed one in the middle)
+      expect(picker.value).toBe('t3');
+      expect(screen.queryByLabelText('Take for clip 2')).not.toBeInTheDocument();
+    });
+
+    it('switching takes changes the clip shown and its status', async () => {
+      const user = userEvent.setup();
+      vi.mocked(getJobsList).mockResolvedValue([
+        wolf(1, { id: 't1', created_at: 1, status: 'completed', prompt: '[Shot 1] Wide shot. First take.' }),
+        wolf(1, { id: 't2', created_at: 2, status: 'failed', prompt: '[Shot 1] Wide shot. Second take.' }),
+      ]);
+      render(<ScenesPage />);
+      await screen.findByRole('heading', { name: 'wolf-lunch' });
+      expect(screen.getByText('First take.')).toBeInTheDocument();
+
+      await user.selectOptions(screen.getByLabelText('Take for clip 1'), 't2');
+      expect(await screen.findByText('Second take.')).toBeInTheDocument();
+      expect(screen.queryByText('First take.')).not.toBeInTheDocument();
+      expect(screen.getAllByText('1 failed').length).toBeGreaterThan(0);
+    });
+  });
+
+  describe('sticky scene header', () => {
+    type Callback = (entries: { isIntersecting: boolean }[]) => void;
+    let callback: Callback | null = null;
+    const disconnect = vi.fn();
+
+    beforeEach(() => {
+      callback = null;
+      disconnect.mockClear();
+      vi.stubGlobal(
+        'IntersectionObserver',
+        class {
+          constructor(cb: Callback) {
+            callback = cb;
+          }
+          observe() {}
+          disconnect = disconnect;
+        },
+      );
+    });
+    afterEach(() => vi.unstubAllGlobals());
+
+    const header = () => screen.getByRole('heading', { name: 'wolf-lunch' }).closest('[data-stuck]') as HTMLElement;
+
+    it('is sticky to the top of the scrolling area', async () => {
+      vi.mocked(getJobsList).mockResolvedValue([wolf(1), wolf(2)]);
+      render(<ScenesPage />);
+      await screen.findByRole('heading', { name: 'wolf-lunch' });
+
+      expect(header()).toHaveClass('sticky', 'top-0');
+      // the timeline bar lives in the sticky block, so it follows the scroll
+      expect(within(header()).getByRole('group', { name: 'Scene timeline' })).toBeInTheDocument();
+    });
+
+    it('adds a shadow only once it has stuck', async () => {
+      vi.mocked(getJobsList).mockResolvedValue([wolf(1), wolf(2)]);
+      render(<ScenesPage />);
+      await screen.findByRole('heading', { name: 'wolf-lunch' });
+
+      expect(header()).toHaveAttribute('data-stuck', 'false');
+      expect(header()).not.toHaveClass('shadow-md');
+
+      // the sentinel above the header has scrolled out of view
+      act(() => callback?.([{ isIntersecting: false }]));
+      expect(header()).toHaveAttribute('data-stuck', 'true');
+      expect(header()).toHaveClass('shadow-md');
+
+      // ...and back into view at the top of the page
+      act(() => callback?.([{ isIntersecting: true }]));
+      expect(header()).toHaveAttribute('data-stuck', 'false');
+      expect(header()).not.toHaveClass('shadow-md');
+    });
+
+    it('stops observing when the scene board goes away', async () => {
+      vi.mocked(getJobsList).mockResolvedValue([wolf(1)]);
+      const { unmount } = render(<ScenesPage />);
+      await screen.findByRole('heading', { name: 'wolf-lunch' });
+      unmount();
+      expect(disconnect).toHaveBeenCalled();
+    });
+  });
+});
diff --git a/apps/fastvideo_studio/src/app/scenes/page.tsx b/apps/fastvideo_studio/src/app/scenes/page.tsx
new file mode 100644
index 000000000..652a6bec0
--- /dev/null
+++ b/apps/fastvideo_studio/src/app/scenes/page.tsx
@@ -0,0 +1,132 @@
+'use client';
+
+import { AlertTriangle, Loader2, RefreshCw } from 'lucide-react';
+import * as React from 'react';
+
+import CreateJobModal from '@/components/jobs/CreateJobModal';
+import { EDITABLE_STATUSES, SceneBoard } from '@/components/scenes/SceneBoard';
+import { Button } from '@/components/ui/button';
+import { Card } from '@/components/ui/card';
+import { NativeSelect } from '@/components/ui/native-select';
+import { getJobsList } from '@/lib/api';
+import { buildScenes } from '@/lib/scenes';
+import type { Job } from '@/lib/types';
+
+export default function ScenesPage() {
+  const [jobs, setJobs] = React.useState<Job[]>([]);
+  const [isLoading, setIsLoading] = React.useState(true);
+  const [error, setError] = React.useState<string | null>(null);
+  const [reloadKey, setReloadKey] = React.useState(0);
+  const [selectedKey, setSelectedKey] = React.useState<string | null>(null);
+  const [openJob, setOpenJob] = React.useState<Job | null>(null);
+  // which re-run of a clip number the user picked, by clip
+  const [choices, setChoices] = React.useState<Record<string, string>>({});
+
+  React.useEffect(() => {
+    let cancelled = false;
+    async function load() {
+      try {
+        const list = await getJobsList('inference');
+        if (cancelled) return;
+        setJobs(list);
+        setError(null);
+      } catch (e) {
+        if (!cancelled) setError(e instanceof Error ? e.message : 'Failed to load jobs');
+      } finally {
+        if (!cancelled) setIsLoading(false);
+      }
+    }
+    void load();
+    return () => {
+      cancelled = true;
+    };
+  }, [reloadKey]);
+
+  function reload() {
+    setIsLoading(true);
+    setReloadKey((k) => k + 1);
+  }
+
+  const scenes = React.useMemo(() => buildScenes(jobs, choices), [jobs, choices]);
+  const scene = scenes.find((s) => s.key === selectedKey) ?? scenes[0];
+
+  return (
+    <div className="mx-auto w-full max-w-[1200px] px-4 pb-12 pt-4">
+      <Card className="p-6">
+        <div className="mb-1 flex flex-wrap items-center gap-3">
+          <h2 className="text-2xl font-semibold text-foreground">Scenes</h2>
+          <Button
+            type="button"
+            variant="outline"
+            size="sm"
+            className="ml-auto gap-1.5"
+            onClick={reload}
+            disabled={isLoading}
+          >
+            <RefreshCw className={isLoading ? 'size-3.5 animate-spin' : 'size-3.5'} aria-hidden />
+            Refresh
+          </Button>
+        </div>
+        <p className="mb-6 text-sm text-muted-foreground">
+          Every shot across your jobs, in order. Jobs named with a shared stem and a number, like{' '}
+          <code className="font-mono text-xs">wolf-lunch-clip-01</code>,{' '}
+          <code className="font-mono text-xs">wolf-lunch-clip-02</code>, are grouped into one scene.
+        </p>
+
+        {isLoading && jobs.length === 0 ? (
+          <div className="flex items-center gap-3 p-8 text-muted-foreground">
+            <Loader2 className="h-6 w-6 animate-spin text-primary" />
+            <span>Loading scenes…</span>
+          </div>
+        ) : error && jobs.length === 0 ? (
+          <div role="alert" className="flex flex-col items-center gap-3 py-8 text-center">
+            <AlertTriangle className="size-6 text-destructive" aria-hidden />
+            <p className="max-w-md text-sm text-muted-foreground">{error}</p>
+            <Button type="button" variant="outline" onClick={reload}>
+              Try Again
+            </Button>
+          </div>
+        ) : !scene ? (
+          <p className="py-8 text-center text-muted-foreground">
+            No inference jobs yet. Create some and they&apos;ll show up here.
+          </p>
+        ) : (
+          <div className="flex flex-col gap-6">
+            {scenes.length > 1 && (
+              <label className="flex max-w-md flex-col gap-1 text-xs text-muted-foreground">
+                Scene
+                <NativeSelect value={scene.key} onChange={(e) => setSelectedKey(e.target.value)}>
+                  {scenes.map((s) => (
+                    <option key={s.key} value={s.key}>
+                      {s.title} ({s.clips.length} clip{s.clips.length === 1 ? '' : 's'})
+                    </option>
+                  ))}
+                </NativeSelect>
+              </label>
+            )}
+            <SceneBoard
+              scene={scene}
+              onOpenJob={setOpenJob}
+              onChooseTake={(takeKey, jobId) => setChoices((c) => ({ ...c, [takeKey]: jobId }))}
+            />
+          </div>
+        )}
+      </Card>
+
+      {openJob && (
+        <CreateJobModal
+          isOpen
+          readOnly={!EDITABLE_STATUSES.includes(openJob.status)}
+          editingJob={openJob}
+          jobType={(openJob.job_type ?? 'inference') as never}
+          workloadType={openJob.workload_type ?? 't2v'}
+          onClose={() => setOpenJob(null)}
+          onSuccess={() => {
+            setOpenJob(null);
+            reload();
+          }}
+        />
+      )}
+    </div>
+  );
+}
diff --git a/apps/fastvideo_studio/src/components/jobs/CreateJobButton.test.tsx b/apps/fastvideo_studio/src/components/jobs/CreateJobButton.test.tsx
index a1636ce3f..bd942756d 100644
--- a/apps/fastvideo_studio/src/components/jobs/CreateJobButton.test.tsx
+++ b/apps/fastvideo_studio/src/components/jobs/CreateJobButton.test.tsx
@@ -1,19 +1,29 @@
 import { render, screen } from '@testing-library/react';
 import userEvent from '@testing-library/user-event';
+import { toast } from 'sonner';
 import { describe, expect, it, vi } from 'vitest';
 
 import CreateJobButton from './CreateJobButton';
 
+vi.mock('sonner', () => ({ toast: { error: vi.fn(), warning: vi.fn() } }));
+
 vi.mock('./CreateJobModal', () => ({
   default: ({
     isOpen,
     workloadType,
+    prefillJob,
   }: {
     isOpen: boolean;
     workloadType: string;
+    prefillJob?: { name?: string; prompt: string } | null;
   }) =>
     isOpen ? (
-      <div role="dialog" data-workload-type={workloadType}>
+      <div
+        role="dialog"
+        data-workload-type={workloadType}
+        data-prefill-name={prefillJob?.name ?? ''}
+        data-prefill-prompt={prefillJob?.prompt ?? ''}
+      >
         Create job form
       </div>
     ) : null,
@@ -27,7 +37,11 @@ describe('CreateJobButton', () => {
     await user.click(screen.getByRole('button', { name: 'Create Job' }));
     await user.click(screen.getByRole('menuitem', { name: /I2V/i }));
 
-    expect(screen.getByRole('dialog')).toHaveAttribute(
+    // The dialog opens a frame after the menu selection (see CreateJobButton's
+    // openModal: deferred via requestAnimationFrame to dodge a Radix
+    // DropdownMenu + Dialog pointer-events race), so wait for it rather than
+    // asserting synchronously.
+    expect(await screen.findByRole('dialog')).toHaveAttribute(
       'data-workload-type',
       'i2v',
     );
@@ -45,9 +59,75 @@ describe('CreateJobButton', () => {
     expect(firstItem).toHaveFocus();
     await user.keyboard('{Enter}');
 
-    expect(screen.getByRole('dialog')).toHaveAttribute(
+    expect(await screen.findByRole('dialog')).toHaveAttribute(
       'data-workload-type',
       't2v',
     );
   });
+
+  describe('importing a job file', () => {
+    const file = (contents: string, name = 'job.json') =>
+      new File([contents], name, { type: 'application/json' });
+
+    it('opens the form for the file\'s workload with the job pre-filled', async () => {
+      const user = userEvent.setup();
+      render(<CreateJobButton jobType="inference" />);
+
+      await user.click(screen.getByRole('button', { name: 'Create Job' }));
+      await user.click(screen.getByRole('menuitem', { name: /Import job JSON/ }));
+      await user.upload(
+        screen.getByLabelText('Import job JSON file'),
+        file(
+          JSON.stringify({
+            model_id: 'MiniMaxAI/MiniMax-H3',
+            name: 'from-file',
+            workload_type: 'i2v',
+            prompt: ['summary:', 'Hi.'],
+          }),
+        ),
+      );
+
+      const dialog = await screen.findByRole('dialog');
+      expect(dialog).toHaveAttribute('data-workload-type', 'i2v');
+      expect(dialog).toHaveAttribute('data-prefill-name', 'from-file');
+      expect(dialog).toHaveAttribute('data-prefill-prompt', 'summary:\nHi.');
+    });
+
+    it('reports an unreadable file instead of opening the form', async () => {
+      const user = userEvent.setup();
+      render(<CreateJobButton jobType="inference" />);
+
+      await user.click(screen.getByRole('button', { name: 'Create Job' }));
+      await user.click(screen.getByRole('menuitem', { name: /Import job JSON/ }));
+      await user.upload(screen.getByLabelText('Import job JSON file'), file('not json', 'bad.json'));
+
+      await vi.waitFor(() => expect(toast.error).toHaveBeenCalled());
+      expect(vi.mocked(toast.error).mock.calls.at(-1)?.[0]).toContain("bad.json");
+      expect(screen.queryByRole('dialog')).not.toBeInTheDocument();
+    });
+
+    it('a normal open after an import starts clean', async () => {
+      const user = userEvent.setup();
+      render(<CreateJobButton jobType="inference" />);
+
+      await user.click(screen.getByRole('button', { name: 'Create Job' }));
+      await user.click(screen.getByRole('menuitem', { name: /Import job JSON/ }));
+      await user.upload(
+        screen.getByLabelText('Import job JSON file'),
+        file(JSON.stringify({ model_id: 'm', name: 'from-file', prompt: 'p' })),
+      );
+      expect(await screen.findByRole('dialog')).toHaveAttribute('data-prefill-name', 'from-file');
+
+      await user.click(screen.getByRole('button', { name: 'Create Job' }));
+      await user.click(await screen.findByRole('menuitem', { name: /T2V/i }));
+      await vi.waitFor(() => expect(screen.getByRole('dialog')).toHaveAttribute('data-prefill-name', ''));
+    });
+
+    it('is not offered on the training pages', async () => {
+      const user = userEvent.setup();
+      render(<CreateJobButton jobType="finetuning" />);
+      await user.click(screen.getByRole('button', { name: 'Create Job' }));
+      expect(screen.queryByRole('menuitem', { name: /Import job JSON/ })).not.toBeInTheDocument();
+    });
+  });
 });
diff --git a/apps/fastvideo_studio/src/components/jobs/CreateJobButton.tsx b/apps/fastvideo_studio/src/components/jobs/CreateJobButton.tsx
index 995423030..e2bd3b4b7 100644
--- a/apps/fastvideo_studio/src/components/jobs/CreateJobButton.tsx
+++ b/apps/fastvideo_studio/src/components/jobs/CreateJobButton.tsx
@@ -6,7 +6,10 @@ import * as DropdownMenu from '@radix-ui/react-dropdown-menu';
 
 import CreateJobModal from '@/components/jobs/CreateJobModal';
 import { Button } from '@/components/ui/button';
+import { toast } from 'sonner';
 import { WORKLOAD_OPTIONS } from '@/lib/jobConfig';
+import { parseJobJson } from '@/lib/jobJson';
+import type { JobLike } from '@/lib/jobToFields';
 import type { JobType } from '@/lib/types';
 import { triggerRefresh } from '@/stores/jobsRefresh';
 
@@ -22,9 +25,32 @@ export default function CreateJobButton({ jobType }: CreateJobButtonProps) {
     options[0]?.type ?? 't2v',
   );
 
-  function openModal(type: string) {
+  // An imported job file pre-fills the form; a normal open starts clean.
+  const [prefillJob, setPrefillJob] = React.useState<JobLike | null>(null);
+  const fileInputRef = React.useRef<HTMLInputElement>(null);
+
+  async function handleImportFile(e: React.ChangeEvent<HTMLInputElement>) {
+    const file = e.target.files?.[0];
+    e.target.value = ''; // allow re-picking the same file
+    if (!file) return;
+    const result = parseJobJson(await file.text());
+    if (!result.ok) {
+      toast.error(`Couldn't import ${file.name}: ${result.error}`);
+      return;
+    }
+    openModal(result.workloadType, result.job);
+  }
+
+  function openModal(type: string, prefill: JobLike | null = null) {
     setWorkloadType(type);
-    setModalOpen(true);
+    setPrefillJob(prefill);
+    // Opening the Dialog synchronously inside the DropdownMenu.Item's
+    // onSelect races the two Radix layers' pointer-events body lock/unlock:
+    // the dropdown's own close cleanup can lose that race, leaving <body>
+    // stuck with `pointer-events: none` after this dialog later closes --
+    // silently blocking every click, including this same button. Deferring
+    // to the next frame lets the dropdown finish closing first.
+    requestAnimationFrame(() => setModalOpen(true));
   }
 
   function handleSuccess() {
@@ -60,15 +86,39 @@ export default function CreateJobButton({ jobType }: CreateJobButtonProps) {
                 </span>
               </DropdownMenu.Item>
             ))}
+            {jobType === 'inference' && (
+              <>
+                <DropdownMenu.Separator className="my-1 h-px bg-border" />
+                <DropdownMenu.Item
+                  onSelect={() => fileInputRef.current?.click()}
+                  className="flex min-h-11 cursor-pointer select-none flex-col justify-center px-4 py-2 text-left text-sm font-medium outline-none data-[highlighted]:bg-secondary"
+                >
+                  Import job JSON…
+                  <span className="mt-0.5 block text-xs font-normal text-muted-foreground">
+                    Fill the form from a job file
+                  </span>
+                </DropdownMenu.Item>
+              </>
+            )}
           </DropdownMenu.Content>
         </DropdownMenu.Portal>
       </DropdownMenu.Root>
+      <input
+        ref={fileInputRef}
+        type="file"
+        accept=".json,application/json"
+        aria-label="Import job JSON file"
+        tabIndex={-1}
+        className="sr-only"
+        onChange={handleImportFile}
+      />
       <CreateJobModal
         isOpen={modalOpen}
         onClose={() => setModalOpen(false)}
         onSuccess={handleSuccess}
         jobType={jobType}
         workloadType={workloadType}
+        prefillJob={prefillJob}
       />
     </>
   );
diff --git a/apps/fastvideo_studio/src/components/jobs/CreateJobModal.test.tsx b/apps/fastvideo_studio/src/components/jobs/CreateJobModal.test.tsx
index a2ae5517d..2b7417a32 100644
--- a/apps/fastvideo_studio/src/components/jobs/CreateJobModal.test.tsx
+++ b/apps/fastvideo_studio/src/components/jobs/CreateJobModal.test.tsx
@@ -4,10 +4,13 @@ import userEvent from '@testing-library/user-event';
 import { beforeEach, describe, expect, it, vi } from 'vitest';
 
 import CreateJobModal from './CreateJobModal';
+import { toast } from 'sonner';
 import { createJob, getDatasets, getModels, uploadImage } from '@/lib/api';
 import { defaultOptionsStore } from '@/stores/defaultOptions';
 import { DEFAULT_OPTIONS } from '@/lib/defaultOptions';
 
+vi.mock('sonner', () => ({ toast: { warning: vi.fn(), error: vi.fn(), success: vi.fn() } }));
+
 vi.mock('@/lib/api', () => ({
   createJob: vi.fn(),
   getModels: vi.fn(),
@@ -206,4 +209,179 @@ describe('CreateJobModal', () => {
 
     await waitFor(() => expect(onSuccess).toHaveBeenCalledTimes(1));
   });
+
+  it('shows the H3 speech-tag console only for an H3 model and inserts at the prompt cursor', async () => {
+    vi.mocked(getModels).mockResolvedValue([
+      ...MODELS,
+      { id: 'minimax-h3/i2v', label: 'MiniMax H3', type: 'i2v' },
+    ]);
+    const user = userEvent.setup();
+    renderModal({ workloadType: 'i2v' });
+
+    await screen.findByRole('option', { name: 'Wan T2V (wan/t2v-1.3b)' });
+    expect(screen.queryByText(/Speech tags/)).not.toBeInTheDocument();
+
+    await user.selectOptions(
+      screen.getByLabelText('Model'),
+      'minimax-h3/i2v',
+    );
+    expect(await screen.findByText(/Speech tags/)).toBeInTheDocument();
+
+    const promptField = screen.getByLabelText('Prompt') as HTMLTextAreaElement;
+    await user.type(promptField, 'Hello there.');
+    promptField.setSelectionRange(5, 5);
+
+    await user.click(screen.getByRole('button', { name: '<pause>' }));
+
+    expect(promptField.value).toBe('Hello <pause> there.');
+    // Focus and cursor return to the prompt field so tags can be chained.
+    await waitFor(() => expect(promptField).toHaveFocus());
+  });
+});
+
+const H3 = { id: 'minimax-h3/i2v', label: 'MiniMax H3', type: 'i2v' };
+
+const BASE_PROMPT = [
+  'For the target video, <Picture 1> is fully referenced.',
+  '',
+  'integrated_multimodal_description: [Shot 1] A close-up. He (S1) says: <d>[English] Sorry.</d> <cutoff>',
+  '',
+  'overall_soundscape: Room tone.',
+  '',
+  'non_diegetic_music: N/A',
+].join('\n');
+
+describe('CreateJobModal with a prefilled job', () => {
+  beforeEach(() => {
+    vi.mocked(getModels).mockResolvedValue([...MODELS, H3]);
+    vi.mocked(toast.warning).mockClear();
+  });
+
+  it('fills the form from the job and creates a new job on submit', async () => {
+    const user = userEvent.setup();
+    const { onSuccess } = renderModal({
+      workloadType: 'i2v',
+      prefillJob: {
+        id: 'imported-1',
+        model_id: H3.id,
+        name: 'from-file',
+        prompt: BASE_PROMPT,
+        workload_type: 'i2v',
+        references: [{ source: '/data/inputs/face.png', media_type: 'image' }],
+        num_frames: 243,
+        seed: 7,
+      },
+    });
+
+    await screen.findByRole('option', { name: /MiniMax H3/ });
+    await waitFor(() => expect(screen.getByLabelText('Model')).toHaveValue(H3.id));
+    expect(screen.getByLabelText('Name (optional)')).toHaveValue('from-file');
+    expect(screen.getByLabelText('Prompt')).toHaveValue(BASE_PROMPT);
+    expect(screen.getByText('face.png')).toBeInTheDocument();
+    expect(screen.getByText('New Inference Job (I2V)')).toBeInTheDocument();
+
+    await user.click(screen.getByRole('button', { name: 'Create Job' }));
+    await waitFor(() => expect(onSuccess).toHaveBeenCalledTimes(1));
+    const payload = vi.mocked(createJob).mock.calls[0][0] as unknown as Record<string, unknown>;
+    expect(payload).toMatchObject({
+      model_id: H3.id,
+      name: 'from-file',
+      prompt: BASE_PROMPT,
+      num_frames: 243,
+      seed: 7,
+      references: [{ source: '/data/inputs/face.png', media_type: 'image' }],
+    });
+  });
+
+  it("warns when the file's model isn't offered for the workload", async () => {
+    renderModal({
+      workloadType: 't2v',
+      prefillJob: { id: 'imported-2', model_id: 'not/a-real-model', prompt: 'p' },
+    });
+    await waitFor(() => expect(toast.warning).toHaveBeenCalled());
+    expect(vi.mocked(toast.warning).mock.calls[0][0]).toContain('not/a-real-model');
+    expect(screen.getByLabelText('Model')).toHaveValue(MODELS[0].id);
+  });
+});
+
+describe('CreateJobModal shot list on a raw prompt', () => {
+  beforeEach(() => {
+    vi.mocked(getModels).mockResolvedValue([...MODELS, H3]);
+  });
+
+  async function openShotList(user: ReturnType<typeof userEvent.setup>) {
+    await user.click(screen.getByRole('button', { name: 'Open shot list' }));
+    return screen.findByRole('dialog', { name: 'Shot list' });
+  }
+
+  it('edits dialogue inside integrated_multimodal_description and leaves the rest of the prompt alone', async () => {
+    const user = userEvent.setup();
+    renderModal({
+      workloadType: 'i2v',
+      prefillJob: {
+        id: 'imported-3',
+        model_id: H3.id,
+        prompt: BASE_PROMPT,
+        references: [{ source: '/data/face.png', media_type: 'image' }],
+      },
+    });
+    await waitFor(() => expect(screen.getByLabelText('Model')).toHaveValue(H3.id));
+
+    const dialog = await openShotList(user);
+    expect(within(dialog).getAllByText(/integrated_multimodal_description/).length).toBeGreaterThan(0);
+    expect(within(dialog).queryByText('Retention analysis')).not.toBeInTheDocument();
+
+    const line = within(dialog).getByPlaceholderText('What the speaker says') as HTMLTextAreaElement;
+    expect(line.value).toBe('Sorry.');
+    await user.type(line, ' Really.');
+    await user.click(within(dialog).getByRole('button', { name: 'Apply' }));
+
+    await waitFor(() =>
+      expect(screen.getByLabelText('Prompt')).toHaveValue(BASE_PROMPT.replace('Sorry.', 'Sorry. Really.')),
+    );
+  });
+
+  it('treats a plain prompt as the shot text and writes shot markers back over it', async () => {
+    const user = userEvent.setup();
+    renderModal({ workloadType: 't2v' });
+    await screen.findByRole('option', { name: /MiniMax H3/ });
+    await user.selectOptions(screen.getByLabelText('Model'), H3.id);
+    await user.type(screen.getByLabelText('Prompt'), 'A dock at dawn.');
+
+    const dialog = await openShotList(user);
+    await user.click(within(dialog).getByRole('button', { name: 'Apply' }));
+    await waitFor(() => expect(screen.getByLabelText('Prompt')).toHaveValue('[Shot 1] A dock at dawn.'));
+  });
+
+  it('keeps retention analysis in step when the raw prompt has a detailed_description and retention_analysis', async () => {
+    const user = userEvent.setup();
+    const sections = [
+      'subject_definitions:',
+      '<Subject 1> is the courier.',
+      '',
+      'retention_analysis:',
+      '<Subject 1> (appears in [Shot 1]): fully_preserved - jacket.',
+      '',
+      'detailed_description:',
+      '[Shot 1] Wide shot. A rooftop with <Subject 1>.',
+    ].join('\n');
+    renderModal({ workloadType: 't2v', prefillJob: { id: 'imported-4', model_id: H3.id, prompt: sections } });
+    await waitFor(() => expect(screen.getByLabelText('Model')).toHaveValue(H3.id));
+
+    const dialog = await openShotList(user);
+    expect(within(dialog).getByText('Retention analysis')).toBeInTheDocument();
+    await user.click(within(dialog).getAllByRole('button', { name: /Add shot/ })[0]);
+    await user.click(within(dialog).getAllByRole('button', { name: '<Subject 1>' })[1]);
+    await user.click(within(dialog).getByRole('button', { name: 'Apply' }));
+
+    const out = (screen.getByLabelText('Prompt') as HTMLTextAreaElement).value;
+    expect(out).toContain('<Subject 1> (appears in [Shot 1], [Shot 2]): fully_preserved - jacket.');
+    expect(out).toContain('subject_definitions:\n<Subject 1> is the courier.');
+  });
+
+  it('is not offered for a non-H3 model', async () => {
+    renderModal({ workloadType: 't2v' });
+    await screen.findByRole('option', { name: 'Wan T2V (wan/t2v-1.3b)' });
+    expect(screen.queryByRole('button', { name: 'Open shot list' })).not.toBeInTheDocument();
+  });
 });
diff --git a/apps/fastvideo_studio/src/components/jobs/CreateJobModal.tsx b/apps/fastvideo_studio/src/components/jobs/CreateJobModal.tsx
index 77b7233ad..bc5afb94a 100644
--- a/apps/fastvideo_studio/src/components/jobs/CreateJobModal.tsx
+++ b/apps/fastvideo_studio/src/components/jobs/CreateJobModal.tsx
@@ -49,8 +49,18 @@ import {
   parseH3Prompt,
   serializeH3Prompt,
   type H3PromptFields,
+  type H3PromptSection,
 } from '@/lib/h3Prompt';
+import { applyH3Tag, type H3Tag } from '@/lib/h3Tags';
+import { H3TagConsole } from '@/components/jobs/H3TagConsole';
+import { H3DialogueEditor } from '@/components/jobs/H3DialogueEditor';
+import { H3ShotListEditor } from '@/components/jobs/H3ShotListEditor';
 import { jobToFormFields, type JobLike } from '@/lib/jobToFields';
+import { readField, shotTarget, writeField, writeShotTarget } from '@/lib/h3PromptFields';
+import { toast } from 'sonner';
+
+/** Prompt fields the speech-tag console can insert into. */
+type TagTarget = 'prompt' | H3PromptSection;
 
 export interface CreateJobModalProps {
   isOpen: boolean;
@@ -60,6 +70,8 @@ export interface CreateJobModalProps {
   workloadType: string;
   /** When set, the modal edits this pending job instead of creating a new one. */
   editingJob?: JobLike | null;
+  /** Seed the form from this job (an imported job file) but still create a new job on submit. */
+  prefillJob?: JobLike | null;
   /** Show the configuration without allowing changes (started/finished jobs). */
   readOnly?: boolean;
 }
@@ -71,6 +83,7 @@ export default function CreateJobModal({
   jobType,
   workloadType,
   editingJob,
+  prefillJob,
   readOnly = false,
 }: CreateJobModalProps) {
   const { options } = useStore(defaultOptionsStore);
@@ -121,11 +134,85 @@ export default function CreateJobModal({
   // H3 is the only registered model with an end frame or references.
   const supportsLastImage = modelId.toLowerCase().includes('minimax-h3');
   const usingReferences = supportsLastImage && references.length > 0;
+  // Same check, named for where it gates H3-only prompt features (speech tags)
+  // rather than the end-frame/reference upload UI.
+  const isH3Model = supportsLastImage;
+
+  // Which prompt field the speech-tag console inserts into: whichever one is
+  // actually visible, tracked via each field's onFocus below.
+  const [activeTagField, setActiveTagField] = React.useState<TagTarget>('prompt');
+  const tagFieldRefs = React.useRef<Partial<Record<TagTarget, HTMLTextAreaElement | null>>>({});
+  const showGuidedFields = usingReferences && useGuidedPrompt;
+  React.useEffect(() => {
+    setActiveTagField(showGuidedFields ? 'detailed_description' : 'prompt');
+  }, [showGuidedFields]);
+
+  function getTagFieldValue(field: TagTarget): string {
+    return field === 'prompt' ? prompt : promptFields[field];
+  }
+  function setTagFieldValue(field: TagTarget, value: string) {
+    if (field === 'prompt') setPrompt(value);
+    else setPromptField(field, value);
+  }
+
+  function insertH3Tag(tag: H3Tag) {
+    const field = activeTagField;
+    const el = tagFieldRefs.current[field] ?? null;
+    const currentValue = getTagFieldValue(field);
+    const start = el?.selectionStart ?? currentValue.length;
+    const end = el?.selectionEnd ?? currentValue.length;
+    const result = applyH3Tag(currentValue, start, end, tag);
+    setTagFieldValue(field, result.value);
+    // The textarea re-renders with the new value before this runs; restore
+    // focus and the cursor position so tags can be chained without reaching
+    // for the mouse each time.
+    requestAnimationFrame(() => {
+      const node = tagFieldRefs.current[field];
+      if (node) {
+        node.focus();
+        node.setSelectionRange(result.start, result.end);
+      }
+    });
+  }
+
+  // Full-screen dialogue editor: opens with a copy of whichever field the
+  // speech-tag console would currently insert into, and on Apply overwrites
+  // that same field wholesale (not just an insertion at the cursor).
+  const [dialogueEditorOpen, setDialogueEditorOpen] = React.useState(false);
+  const dialogueEditorField = activeTagField;
+
+  // Shot list: writes into detailed_description AND retention_analysis
+  // together, so unlike the speech-tag console/dialogue editor it isn't
+  // scoped to "whichever field is focused" -- it's specific to those two.
+  const [shotListOpen, setShotListOpen] = React.useState(false);
+  // 'sections' edits the guided Detailed description; 'prompt' edits the shots
+  // inside the raw prompt (base format, or a plain prompt).
+  const [shotListMode, setShotListMode] = React.useState<'sections' | 'prompt'>('sections');
+  function applyShotList(detailedDescription: string, retentionAnalysis: string) {
+    setPromptFields((prev) => ({
+      ...prev,
+      detailed_description: detailedDescription,
+      retention_analysis: retentionAnalysis,
+    }));
+  }
+
+  const promptShots = React.useMemo(() => shotTarget(prompt), [prompt]);
+  const promptSubjects = React.useMemo(() => readField(prompt, 'subject_definitions') ?? '', [prompt]);
+  const promptRetention = React.useMemo(() => readField(prompt, 'retention_analysis'), [prompt]);
+  const promptSyncsRetention = promptShots.field === 'detailed_description' && promptRetention !== null;
+
+  function applyShotListToPrompt(shotText: string, retentionAnalysis: string) {
+    let next = writeShotTarget(prompt, promptShots, shotText);
+    if (promptSyncsRetention) next = writeField(next, 'retention_analysis', retentionAnalysis);
+    setPrompt(next);
+  }
 
   // JobCard re-renders on every job-list poll, so `editingJob` is a fresh
   // object each time. Effects must depend on these, never on the object.
-  const editingJobId = editingJob?.id ?? null;
-  const editingJobModelId = editingJob?.model_id ?? null;
+  const seedJob = editingJob ?? prefillJob ?? null;
+  const seedJobId = seedJob?.id ?? null;
+  const seedJobModelId = seedJob?.model_id ?? null;
+  const isPrefill = !editingJob && !!prefillJob;
 
   // Layerwise offload and FSDP compete for the DiT weights and FastVideoArgs
   // silently picks a winner (fastvideo_args.py:859); resolve it visibly here.
@@ -204,10 +291,10 @@ export default function CreateJobModal({
     const justOpened = isOpen && !justOpenedRef.current;
     justOpenedRef.current = isOpen;
     if (!justOpened) return;
-    if (editingJob) {
+    if (seedJob) {
       // Must not fall through to the defaults below: a partially-seeded
       // form silently edits values the user never saw.
-      const f = jobToFormFields(editingJob);
+      const f = jobToFormFields(seedJob);
       setModelId(f.modelId);
       setName(f.name);
       setPrompt(f.prompt);
@@ -296,7 +383,7 @@ export default function CreateJobModal({
       setRealScoreModelPath('');
       setFakeScoreModelPath('');
     }
-  }, [isOpen, workloadType, inferenceWorkload, options, editingJobId]);
+  }, [isOpen, workloadType, inferenceWorkload, options, seedJobId]);
 
   // Load the models available for this workload.
   React.useEffect(() => {
@@ -319,7 +406,7 @@ export default function CreateJobModal({
         // When editing, the job's own model wins over the workload default --
         // this resolves after the seeding effect, so choosing a default here
         // would silently swap the model out from under the user.
-        const editedId = editingJobModelId;
+        const editedId = seedJobModelId;
         const chosen =
           editedId && ids.includes(editedId)
             ? editedId
@@ -327,6 +414,11 @@ export default function CreateJobModal({
               ? defaultId
               : (list[0]?.id ?? '');
         setModelId(chosen);
+        if (isPrefill && editedId && chosen !== editedId) {
+          toast.warning(
+            `${editedId} isn't available for ${inferenceWorkload.toUpperCase()}${chosen ? `, so ${chosen} is selected instead` : ''}.`,
+          );
+        }
         if (workloadType === 'dmd_t2v') {
           setRealScoreModelPath(chosen);
           setFakeScoreModelPath(chosen);
@@ -347,7 +439,7 @@ export default function CreateJobModal({
     return () => {
       stale = true;
     };
-  }, [isOpen, inferenceWorkload, workloadType, editingJobModelId]);
+  }, [isOpen, inferenceWorkload, workloadType, seedJobModelId, isPrefill]);
 
   // Training jobs need a dataset; load the ready datasets when relevant.
   React.useEffect(() => {
@@ -857,6 +949,22 @@ export default function CreateJobModal({
             </FieldRow>
           )}
 
+          {isH3Model && isInference && workloadType !== 't2i' && (
+            <div className="flex flex-col gap-2">
+              <H3TagConsole onInsert={insertH3Tag} disabled={isSubmitting} />
+              <Button
+                type="button"
+                variant="outline"
+                size="sm"
+                className="self-start"
+                onClick={() => setDialogueEditorOpen(true)}
+                disabled={isSubmitting}
+              >
+                Open dialogue editor
+              </Button>
+            </div>
+          )}
+
           {usingReferences && useGuidedPrompt ? (
             /* Six-section format from the model's reference prompt guide. */
             <>
@@ -868,12 +976,31 @@ export default function CreateJobModal({
                 >
                   <Textarea
                     id={`modal-prompt-${section}`}
+                    ref={(el) => {
+                      tagFieldRefs.current[section] = el;
+                    }}
                     value={promptFields[section]}
                     onChange={(e) => setPromptField(section, e.target.value)}
+                    onFocus={() => setActiveTagField(section)}
                     rows={section === 'detailed_description' ? 5 : 2}
                     placeholder={H3_SECTION_HINTS[section]}
                     disabled={isSubmitting}
                   />
+                  {section === 'detailed_description' && (
+                    <Button
+                      type="button"
+                      variant="outline"
+                      size="sm"
+                      className="self-start"
+                      onClick={() => {
+                        setShotListMode('sections');
+                        setShotListOpen(true);
+                      }}
+                      disabled={isSubmitting}
+                    >
+                      Open shot list
+                    </Button>
+                  )}
                 </FieldRow>
               ))}
               <button
@@ -893,8 +1020,12 @@ export default function CreateJobModal({
               >
                 <Textarea
                   id="modal-prompt"
+                  ref={(el) => {
+                    tagFieldRefs.current.prompt = el;
+                  }}
                   value={prompt}
                   onChange={(e) => setPrompt(e.target.value)}
+                  onFocus={() => setActiveTagField('prompt')}
                   rows={isInference ? 3 : 2}
                   placeholder={
                     isInference
@@ -904,6 +1035,21 @@ export default function CreateJobModal({
                   required={!(usingReferences && useGuidedPrompt)}
                   disabled={isSubmitting}
                 />
+                {isH3Model && isInference && workloadType !== 't2i' && (
+                  <Button
+                    type="button"
+                    variant="outline"
+                    size="sm"
+                    className="self-start"
+                    onClick={() => {
+                      setShotListMode('prompt');
+                      setShotListOpen(true);
+                    }}
+                    disabled={isSubmitting}
+                  >
+                    Open shot list
+                  </Button>
+                )}
               </FieldRow>
               {usingReferences && (
                 <button
@@ -1379,6 +1525,38 @@ export default function CreateJobModal({
             </Button>
           </div>
         </form>
+
+        {/* Nested in the React tree (still portaled to <body>) so Radix treats
+            clicks and focus inside these as inside this dialog. As siblings
+            they counted as "outside", and closing one could close this too. */}
+        <H3DialogueEditor
+          open={dialogueEditorOpen}
+          initialValue={getTagFieldValue(dialogueEditorField)}
+          onApply={(value) => setTagFieldValue(dialogueEditorField, value)}
+          onClose={() => setDialogueEditorOpen(false)}
+        />
+        <H3ShotListEditor
+          open={shotListOpen}
+          subjectDefinitions={shotListMode === 'sections' ? promptFields.subject_definitions : promptSubjects}
+          initialDetailedDescription={
+            shotListMode === 'sections' ? promptFields.detailed_description : promptShots.body
+          }
+          initialRetentionAnalysis={
+            shotListMode === 'sections' ? promptFields.retention_analysis : (promptRetention ?? '')
+          }
+          targetLabel={
+            shotListMode === 'sections'
+              ? undefined
+              : promptShots.field === 'integrated_multimodal_description'
+                ? 'integrated_multimodal_description'
+                : promptShots.field === 'detailed_description'
+                  ? 'Detailed description'
+                  : 'the prompt'
+          }
+          withRetention={shotListMode === 'sections' || promptSyncsRetention}
+          onApply={shotListMode === 'sections' ? applyShotList : applyShotListToPrompt}
+          onClose={() => setShotListOpen(false)}
+        />
       </DialogContent>
     </Dialog>
   );
diff --git a/apps/fastvideo_studio/src/components/jobs/H3DialogueEditor.tsx b/apps/fastvideo_studio/src/components/jobs/H3DialogueEditor.tsx
new file mode 100644
index 000000000..80cb7a2e0
--- /dev/null
+++ b/apps/fastvideo_studio/src/components/jobs/H3DialogueEditor.tsx
@@ -0,0 +1,346 @@
+'use client';
+
+import * as React from 'react';
+
+import { Button } from '@/components/ui/button';
+import {
+  Dialog,
+  DialogContent,
+  DialogDescription,
+  DialogFooter,
+  DialogHeader,
+  DialogTitle,
+} from '@/components/ui/dialog';
+import { Textarea } from '@/components/ui/textarea';
+import { H3DialoguePreview } from '@/components/jobs/H3DialoguePreview';
+import { H3TagPalette } from '@/components/jobs/H3TagPalette';
+import { useUndoableText } from '@/hooks/useUndoableText';
+import { applyH3Preset, applyH3Tag, type H3Preset, type H3Tag } from '@/lib/h3Tags';
+import {
+  SPEECH_RATES,
+  dialogueTree,
+  estimateDuration,
+  type DialogueNode,
+  type SpeechRate,
+} from '@/lib/h3DialogueEstimate';
+import { cn } from '@/lib/utils';
+
+export interface H3DialogueEditorProps {
+  open: boolean;
+  /** Seeded from whichever prompt field the editor was opened for. */
+  initialValue: string;
+  onApply: (value: string) => void;
+  onClose: () => void;
+}
+
+/** Every `<d>...</d>` block's inner text -- what's actually spoken, ignoring surrounding shot/camera prose. */
+function extractDialogueBlocks(text: string): string[] {
+  const blocks: string[] = [];
+  const re = /<d>([\s\S]*?)<\/d>/g;
+  let m: RegExpExecArray | null;
+  while ((m = re.exec(text))) blocks.push(m[1].replace(/^\s*\[[^\]]*\]\s*/, '')); // drop a leading [Language] marker
+  return blocks;
+}
+
+/**
+ * Full-screen dialogue editor for MiniMax-H3 speech tags: the same tag
+ * registry as `H3TagConsole`, plus a duration estimate (scoped to `<d>`
+ * blocks so scene-description prose doesn't inflate it), a styled
+ * performance preview, and emotion presets. Opens with a copy of the target
+ * field's text; `onApply` hands the edited text back, `onClose` discards it.
+ * The shot list editor reuses the same palette, preview, and undo history.
+ */
+export function H3DialogueEditor({ open, initialValue, onApply, onClose }: H3DialogueEditorProps) {
+  const { text, type, commit, undo, redo, reset, canUndo, canRedo } = useUndoableText(initialValue);
+  const [pace, setPace] = React.useState<SpeechRate>('Natural');
+  const [calib, setCalib] = React.useState(1);
+  const [actualSeconds, setActualSeconds] = React.useState('');
+  const taRef = React.useRef<HTMLTextAreaElement>(null);
+  const pending = React.useRef<[number, number] | null>(null);
+
+  function restoreCaretAtEnd(restored: string | null) {
+    if (restored != null) pending.current = [restored.length, restored.length];
+  }
+
+  // Re-seed from the field's current value each time the editor opens, the
+  // same "seed on the open transition only" guard CreateJobModal itself
+  // uses -- otherwise an in-progress edit here could be clobbered by a
+  // stray parent re-render while open. History resets with it: undo
+  // shouldn't reach back into a previous time this editor was opened.
+  const wasOpenRef = React.useRef(false);
+  React.useEffect(() => {
+    if (open && !wasOpenRef.current) {
+      reset(initialValue);
+      setActualSeconds('');
+      setCalib(1);
+    }
+    wasOpenRef.current = open;
+    // eslint-disable-next-line react-hooks/exhaustive-deps
+  }, [open]);
+
+  React.useEffect(() => {
+    if (pending.current && taRef.current) {
+      const [s, e] = pending.current;
+      taRef.current.focus();
+      taRef.current.setSelectionRange(s, e);
+      pending.current = null;
+    }
+  });
+
+  function handleKeyDown(e: React.KeyboardEvent<HTMLTextAreaElement>) {
+    if (!(e.metaKey || e.ctrlKey) || e.key.toLowerCase() !== 'z') return;
+    e.preventDefault();
+    restoreCaretAtEnd(e.shiftKey ? redo() : undo());
+  }
+
+  function insertTag(tag: H3Tag) {
+    const ta = taRef.current;
+    const result = applyH3Tag(text, ta?.selectionStart ?? text.length, ta?.selectionEnd ?? text.length, tag);
+    commit(result.value);
+    pending.current = [result.start, result.end];
+  }
+
+  function applyPreset(p: H3Preset) {
+    const ta = taRef.current;
+    const result = applyH3Preset(text, ta?.selectionStart ?? text.length, ta?.selectionEnd ?? text.length, p);
+    commit(result.value);
+    pending.current = [result.start, result.end];
+  }
+
+  const dialogueBlocks = React.useMemo(() => extractDialogueBlocks(text), [text]);
+  // Falls back to the whole field only when it contains no <d> blocks at
+  // all yet -- e.g. a fresh field the user is about to write dialogue into.
+  const segments = dialogueBlocks.length > 0 ? dialogueBlocks : [text];
+  const estimates = React.useMemo(
+    () => segments.map((s) => estimateDuration(s, SPEECH_RATES[pace], calib)),
+    // eslint-disable-next-line react-hooks/exhaustive-deps
+    [segments.join('\n'), pace, calib],
+  );
+  const totalSeconds = estimates.reduce((sum, e) => sum + e.total, 0);
+
+  function calibrate() {
+    const actual = parseFloat(actualSeconds);
+    if (!actual || actual <= 0 || segments.length === 0) return;
+    const base = segments.reduce((sum, s) => sum + estimateDuration(s, SPEECH_RATES[pace], 1).total, 0);
+    if (base > 0) setCalib(actual / base);
+  }
+
+  /* ---- rough browser-voice preview, for rhythm/pacing only ---- */
+  const synth = typeof window !== 'undefined' ? window.speechSynthesis : null;
+  const [playing, setPlaying] = React.useState(false);
+  const playTokenRef = React.useRef({ cancelled: true });
+
+  function stopPreview() {
+    playTokenRef.current.cancelled = true;
+    synth?.cancel();
+    setPlaying(false);
+  }
+
+  function playPreview() {
+    if (!synth) return;
+    stopPreview();
+    const token = { cancelled: false };
+    playTokenRef.current = token;
+    setPlaying(true);
+    const VOX: Record<string, string> = {
+      laughs: 'ha ha ha', chuckle: 'heh', sighs: 'haah', gasp: 'ah', uh: 'uh', mhm: 'mhm', phew: 'phew',
+    };
+    const atoms: { say?: string; gap?: number }[] = [];
+    for (const [i, seg] of segments.entries()) {
+      if (i > 0) atoms.push({ gap: 0.6 });
+      const walk = (nodes: DialogueNode[]) => {
+        for (const n of nodes) {
+          if (n.t === 'text') {
+            const t = n.v.replace(/\s+/g, ' ').trim();
+            if (t) atoms.push({ say: t });
+          } else if (n.t === 'point') {
+            atoms.push(VOX[n.name] ? { say: VOX[n.name] } : { gap: 0.25 });
+          } else {
+            if (n.name === 'humming') atoms.push({ say: 'hmm hmm' });
+            else walk(n.children);
+          }
+        }
+      };
+      walk(dialogueTree(seg));
+    }
+    let i = 0;
+    const step = () => {
+      if (token.cancelled) return;
+      if (i >= atoms.length) {
+        setPlaying(false);
+        return;
+      }
+      const a = atoms[i++];
+      if (a.gap != null) {
+        setTimeout(() => { if (!token.cancelled) step(); }, a.gap * 1000);
+        return;
+      }
+      const u = new SpeechSynthesisUtterance(a.say);
+      u.rate = 0.95;
+      u.onend = () => { if (!token.cancelled) step(); };
+      u.onerror = () => { if (!token.cancelled) step(); };
+      synth.speak(u);
+    };
+    step();
+  }
+
+  React.useEffect(() => () => stopPreview(), []); // eslint-disable-line react-hooks/exhaustive-deps
+
+  function handleApply() {
+    onApply(text);
+    onClose();
+  }
+
+  return (
+    <Dialog open={open} onOpenChange={(o) => { if (!o) onClose(); }}>
+      <DialogContent className="flex h-[92vh] max-h-[920px] w-[96vw] max-w-[1180px] flex-col gap-0 overflow-hidden p-0">
+        <DialogHeader className="border-b border-border px-5 py-4">
+          <DialogTitle>Dialogue editor</DialogTitle>
+          <DialogDescription>
+            Click a tag to insert it at the cursor, or select text first to wrap it. Applying writes the edited text
+            back into the field you opened this from.
+          </DialogDescription>
+        </DialogHeader>
+
+        <div className="flex min-h-0 flex-1">
+          {/* palette */}
+          <H3TagPalette
+            onInsertTag={insertTag}
+            onApplyPreset={applyPreset}
+            className="w-64 shrink-0 overflow-y-auto border-r border-border p-3"
+          />
+
+          {/* editor + preview */}
+          <div className="flex min-w-0 flex-1 flex-col gap-4 overflow-y-auto p-4">
+            {/* duration estimate */}
+            <div className="rounded-xl border border-border bg-secondary/30 p-4">
+              <div className="flex flex-wrap items-baseline justify-between gap-3">
+                <div className="flex items-baseline gap-3">
+                  <span className="text-2xl font-semibold tracking-tight text-foreground">
+                    {totalSeconds.toFixed(1)}<span className="ml-1 text-sm font-normal text-muted-foreground">s spoken</span>
+                  </span>
+                  <span className="font-mono text-xs text-muted-foreground">
+                    {dialogueBlocks.length > 0
+                      ? `across ${dialogueBlocks.length} <d> block${dialogueBlocks.length === 1 ? '' : 's'}`
+                      : 'no <d> blocks yet — estimating the whole field'}
+                  </span>
+                </div>
+                <div className="flex items-center gap-1 rounded-lg border border-border bg-card p-1">
+                  {(Object.keys(SPEECH_RATES) as SpeechRate[]).map((p) => (
+                    <button
+                      key={p}
+                      type="button"
+                      onClick={() => setPace(p)}
+                      className={cn(
+                        'rounded-md px-2.5 py-1 text-xs transition-colors',
+                        pace === p ? 'bg-secondary text-foreground' : 'text-muted-foreground hover:bg-accent',
+                      )}
+                    >
+                      {p}
+                    </button>
+                  ))}
+                </div>
+              </div>
+
+              {segments.length > 1 && (
+                <div className="mt-3 flex flex-wrap gap-1.5">
+                  {estimates.map((e, i) => (
+                    <span key={i} className="rounded-md border border-border bg-card px-2 py-0.5 font-mono text-[11px] text-muted-foreground">
+                      block {i + 1}: {e.total.toFixed(1)}s
+                    </span>
+                  ))}
+                </div>
+              )}
+
+              <div className="mt-3 flex flex-wrap items-center gap-2 text-xs text-muted-foreground">
+                <span title="Enter the real length of a clip you already generated, to tune the estimate to your setup">
+                  calibrate
+                </span>
+                <input
+                  value={actualSeconds}
+                  onChange={(e) => setActualSeconds(e.target.value)}
+                  placeholder="actual s"
+                  inputMode="decimal"
+                  className="w-16 rounded-md border border-input bg-card px-2 py-1 text-foreground"
+                />
+                <Button type="button" variant="outline" size="sm" onClick={calibrate}>Set</Button>
+                {Math.abs(calib - 1) > 0.001 && (
+                  <button
+                    type="button"
+                    onClick={() => { setCalib(1); setActualSeconds(''); }}
+                    className="font-mono text-muted-foreground/80 hover:text-foreground"
+                    title="Reset calibration"
+                  >
+                    ×{calib.toFixed(2)}
+                  </button>
+                )}
+              </div>
+            </div>
+
+            {/* audio preview */}
+            {synth && (
+              <div className="flex items-center gap-3 rounded-xl border border-border bg-secondary/30 p-3">
+                <Button type="button" size="sm" onClick={playing ? stopPreview : playPreview}>
+                  {playing ? '■ Stop' : '▶ Play preview'}
+                </Button>
+                <span className="text-xs text-muted-foreground">
+                  A stand-in browser voice for rhythm only — it won&apos;t match MiniMax&apos;s voice.
+                </span>
+              </div>
+            )}
+
+            {/* the field itself */}
+            <div className="flex items-center justify-between">
+              <span className="text-xs font-semibold text-foreground">Text</span>
+              <div className="flex items-center gap-1">
+                <Button
+                  type="button"
+                  variant="outline"
+                  size="sm"
+                  onClick={() => restoreCaretAtEnd(undo())}
+                  disabled={!canUndo}
+                  title="Undo (⌘/Ctrl+Z)"
+                >
+                  ↶ Undo
+                </Button>
+                <Button
+                  type="button"
+                  variant="outline"
+                  size="sm"
+                  onClick={() => restoreCaretAtEnd(redo())}
+                  disabled={!canRedo}
+                  title="Redo (⌘/Ctrl+Shift+Z)"
+                >
+                  ↷ Redo
+                </Button>
+              </div>
+            </div>
+            <Textarea
+              ref={taRef}
+              value={text}
+              onChange={(e) => type(e.target.value)}
+              onKeyDown={handleKeyDown}
+              spellCheck={false}
+              rows={8}
+              className="flex-1 font-mono text-sm"
+            />
+
+            {/* performance preview */}
+            <div>
+              <div className="mb-2 text-xs font-semibold text-foreground">Performance preview</div>
+              <H3DialoguePreview
+                text={text}
+                className="rounded-xl border border-border bg-card p-4 text-sm leading-7 text-foreground"
+              />
+            </div>
+          </div>
+        </div>
+
+        <DialogFooter className="border-t border-border px-5 py-3">
+          <Button type="button" variant="outline" onClick={onClose}>Cancel</Button>
+          <Button type="button" onClick={handleApply}>Apply to field</Button>
+        </DialogFooter>
+      </DialogContent>
+    </Dialog>
+  );
+}
diff --git a/apps/fastvideo_studio/src/components/jobs/H3DialoguePreview.tsx b/apps/fastvideo_studio/src/components/jobs/H3DialoguePreview.tsx
new file mode 100644
index 000000000..6528b64e9
--- /dev/null
+++ b/apps/fastvideo_studio/src/components/jobs/H3DialoguePreview.tsx
@@ -0,0 +1,44 @@
+'use client';
+
+import * as React from 'react';
+
+import { CATEGORY_STYLE } from '@/components/jobs/H3TagPalette';
+import { dialogueTree, type DialogueNode } from '@/lib/h3DialogueEstimate';
+import { H3_TAG_CATEGORIES } from '@/lib/h3Tags';
+import { cn } from '@/lib/utils';
+
+const CATEGORY_BY_TAG_NAME: Record<string, string> = Object.fromEntries(
+  H3_TAG_CATEGORIES.flatMap((c) => c.tags.map((t) => [t.open.replace(/[<>]/g, ''), c.id])),
+);
+
+function renderPreviewNodes(nodes: DialogueNode[], keyPrefix = ''): React.ReactNode[] {
+  return nodes.map((n, i) => {
+    const key = `${keyPrefix}${i}`;
+    if (n.t === 'text') return <span key={key}>{n.v}</span>;
+    if (n.t === 'point') {
+      const cat = CATEGORY_BY_TAG_NAME[n.name] ?? 'body';
+      const style = CATEGORY_STYLE[cat] ?? CATEGORY_STYLE.body;
+      return (
+        <span
+          key={key}
+          className={cn(
+            'mx-0.5 inline-block rounded border px-1.5 py-0.5 font-mono text-[11px] leading-none',
+            style.chip,
+          )}
+        >
+          {n.name}
+        </span>
+      );
+    }
+    const inner = renderPreviewNodes(n.children, `${key}-`);
+    if (n.name === 'i') return <strong key={key} className="italic">{inner}</strong>;
+    if (n.name === 'whisper') return <span key={key} className="text-[0.92em] tracking-wide opacity-60">{inner}</span>;
+    if (n.name === 'humming') return <em key={key} className="underline decoration-wavy decoration-emerald-500 underline-offset-4">{inner}</em>;
+    return <span key={key}>{inner}</span>;
+  });
+}
+
+/** Styled "performance" rendering of speech tags: chips for point tags, styling for wrapping ones. */
+export function H3DialoguePreview({ text, className }: { text: string; className?: string }) {
+  return <div className={className}>{renderPreviewNodes(dialogueTree(text))}</div>;
+}
diff --git a/apps/fastvideo_studio/src/components/jobs/H3Editors.test.tsx b/apps/fastvideo_studio/src/components/jobs/H3Editors.test.tsx
new file mode 100644
index 000000000..475bcf4f6
--- /dev/null
+++ b/apps/fastvideo_studio/src/components/jobs/H3Editors.test.tsx
@@ -0,0 +1,184 @@
+import { render, screen, within } from '@testing-library/react';
+import userEvent from '@testing-library/user-event';
+import { describe, expect, it, vi } from 'vitest';
+
+import { H3DialogueEditor } from './H3DialogueEditor';
+import { H3ShotListEditor } from './H3ShotListEditor';
+
+const SUBJECTS = '<Subject 1> is the courier.\n<Subject 2> is the rival.';
+const DETAILED = [
+  '[Shot 1] Wide shot. A rooftop at night.',
+  '[Shot 2] Close-up shot, handheld. He turns. (S2) <d>[English] I had no choice.</d>',
+].join('\n\n');
+
+describe('H3DialogueEditor (shared palette + undo)', () => {
+  it('inserts a tag at the cursor, undoes it, and applies the edited text', async () => {
+    const user = userEvent.setup();
+    const onApply = vi.fn();
+    render(<H3DialogueEditor open initialValue="Hello there" onApply={onApply} onClose={() => {}} />);
+
+    const textarea = screen.getByDisplayValue('Hello there') as HTMLTextAreaElement;
+    textarea.focus();
+    textarea.setSelectionRange(5, 5);
+    await user.click(screen.getByRole('button', { name: /<pause>/ }));
+    expect(textarea.value).toBe('Hello <pause> there');
+
+    await user.click(screen.getByRole('button', { name: /Undo/ }));
+    expect(textarea.value).toBe('Hello there');
+
+    await user.click(screen.getByRole('button', { name: 'Nervous' }));
+    expect(textarea.value).toContain('<uh> <stutter>');
+
+    await user.click(screen.getByRole('button', { name: 'Apply to field' }));
+    expect(onApply).toHaveBeenCalledWith(expect.stringContaining('<uh> <stutter>'));
+  });
+});
+
+describe('H3ShotListEditor dialogue tooling', () => {
+  function renderEditor(onApply = vi.fn()) {
+    render(
+      <H3ShotListEditor
+        open
+        subjectDefinitions={SUBJECTS}
+        initialDetailedDescription={DETAILED}
+        initialRetentionAnalysis=""
+        onApply={onApply}
+        onClose={() => {}}
+      />,
+    );
+    return onApply;
+  }
+
+  it('disables the palette until a shot dialogue is focused', () => {
+    renderEditor();
+    expect(screen.getByText(/Click into a shot's dialogue/)).toBeInTheDocument();
+    expect(screen.getByRole('button', { name: /<pause>/ })).toBeDisabled();
+  });
+
+  it('inserts a palette tag into the focused shot dialogue and writes it on apply', async () => {
+    const user = userEvent.setup();
+    const onApply = renderEditor();
+
+    const dialogueBoxes = screen.getAllByPlaceholderText('What the speaker says') as HTMLTextAreaElement[];
+    expect(dialogueBoxes).toHaveLength(1);
+    expect(dialogueBoxes[0].value).toBe('I had no choice.');
+
+    await user.click(dialogueBoxes[0]);
+    dialogueBoxes[0].setSelectionRange(1, 1);
+    expect(screen.getByText('Tags insert into Shot 2 dialogue')).toBeInTheDocument();
+
+    await user.click(screen.getByRole('button', { name: /<pause>/ }));
+    expect(dialogueBoxes[0].value).toBe('I <pause> had no choice.');
+
+    await user.click(screen.getByRole('button', { name: 'Apply' }));
+    expect(onApply).toHaveBeenCalledTimes(1);
+    expect(onApply.mock.calls[0][0]).toContain('<d>[English] I <pause> had no choice.</d>');
+  });
+
+  it('adds and removes dialogue lines, writing every line on apply', async () => {
+    const user = userEvent.setup();
+    const onApply = renderEditor();
+
+    await user.click(screen.getAllByRole('button', { name: '+ Add line' })[1]);
+    const boxes = screen.getAllByPlaceholderText('What the speaker says') as HTMLTextAreaElement[];
+    await user.type(boxes[1], 'Then go.');
+    await user.click(screen.getByRole('button', { name: 'Apply' }));
+    expect(onApply.mock.calls[0][0]).toContain(
+      '(S2) <d>[English] I had no choice.</d> (S2) <d>[English] Then go.</d>',
+    );
+  });
+
+  it('shows the styled preview once a shot dialogue contains tags', async () => {
+    const user = userEvent.setup();
+    renderEditor();
+    const boxes = screen.getAllByPlaceholderText('What the speaker says') as HTMLTextAreaElement[];
+    await user.click(boxes[0]);
+    await user.click(screen.getByRole('button', { name: /<long pause>/ }));
+
+    const shot2 = boxes[0].closest('[id^="shot-card-"]') as HTMLElement;
+    expect(within(shot2).getByText('long pause')).toBeInTheDocument();
+  });
+});
+
+describe('H3ShotListEditor ordered body', () => {
+  function renderWith(detailed: string) {
+    const onApply = vi.fn();
+    render(
+      <H3ShotListEditor
+        open
+        subjectDefinitions={SUBJECTS}
+        initialDetailedDescription={detailed}
+        initialRetentionAnalysis=""
+        onApply={onApply}
+        onClose={() => {}}
+      />,
+    );
+    return onApply;
+  }
+
+  it('shows hand-written dialogue as an editable line between its prose and applies the text unchanged', async () => {
+    const user = userEvent.setup();
+    const text = "[Shot 1] He (S1) says: <d>[English] I think you're sorry.</d> <cutoff>";
+    const onApply = renderWith(text);
+
+    expect(screen.getByText(/keeps its original wording/)).toBeInTheDocument();
+    const lines = screen.getAllByPlaceholderText('What the speaker says') as HTMLTextAreaElement[];
+    expect(lines).toHaveLength(1);
+    expect(lines[0].value).toBe("I think you're sorry.");
+    const prose = screen.getAllByPlaceholderText("What's happening in this shot") as HTMLTextAreaElement[];
+    expect(prose.map((p) => p.value)).toEqual(['He (S1) says: ', ' <cutoff>']);
+
+    await user.click(screen.getByRole('button', { name: 'Apply' }));
+    expect(onApply.mock.calls[0][0]).toBe(text);
+  });
+
+  it('editing a hand-written line keeps the tag after it where it was', async () => {
+    const user = userEvent.setup();
+    const onApply = renderWith("[Shot 1] He (S1) says: <d>[English] Sorry.</d> <cutoff>");
+    const [line] = screen.getAllByPlaceholderText('What the speaker says') as HTMLTextAreaElement[];
+    await user.type(line, ' Really.');
+    await user.click(screen.getByRole('button', { name: 'Apply' }));
+    expect(onApply.mock.calls[0][0]).toBe("[Shot 1] He (S1) says: <d>[English] Sorry. Really.</d> <cutoff>");
+  });
+
+  it('reorders dialogue lines with the move buttons', async () => {
+    const user = userEvent.setup();
+    const onApply = renderWith('[Shot 1] Wide shot. (S1) <d>[English] One.</d> (S2) <d>[English] Two.</d>');
+    await user.click(screen.getByRole('button', { name: 'Move line 2 up' }));
+    await user.click(screen.getByRole('button', { name: 'Apply' }));
+    expect(onApply.mock.calls[0][0]).toBe(
+      '[Shot 1] Wide shot. (S2) <d>[English] Two.</d> (S1) <d>[English] One.</d>',
+    );
+  });
+
+  it('adds a text block after the existing content', async () => {
+    const user = userEvent.setup();
+    const onApply = renderWith('[Shot 1] Wide shot. A dock.');
+    await user.click(screen.getByRole('button', { name: '+ Add text' }));
+    const prose = screen.getAllByPlaceholderText("What's happening in this shot") as HTMLTextAreaElement[];
+    expect(prose).toHaveLength(2);
+    await user.type(prose[1], 'Rain falls.');
+    await user.click(screen.getByRole('button', { name: 'Apply' }));
+    expect(onApply.mock.calls[0][0]).toBe('[Shot 1] Wide shot. A dock. Rain falls.');
+  });
+
+  it('removes a text block', async () => {
+    const user = userEvent.setup();
+    const onApply = renderWith('[Shot 1] Wide shot. A dock. (S1) <d>[English] Hi.</d>');
+    await user.click(screen.getByRole('button', { name: 'Remove text 1' }));
+    await user.click(screen.getByRole('button', { name: 'Apply' }));
+    expect(onApply.mock.calls[0][0]).toBe('[Shot 1] Wide shot. (S1) <d>[English] Hi.</d>');
+  });
+
+  it('adds a new line at the end of the shot, after any closing prose', async () => {
+    const user = userEvent.setup();
+    const onApply = renderWith('[Shot 1] Wide shot. He speaks. (S1) <d>[English] First.</d> He pauses.');
+    await user.click(screen.getByRole('button', { name: '+ Add line' }));
+    const lines = screen.getAllByPlaceholderText('What the speaker says') as HTMLTextAreaElement[];
+    await user.type(lines[1], 'Later.');
+    await user.click(screen.getByRole('button', { name: 'Apply' }));
+    expect(onApply.mock.calls[0][0]).toBe(
+      '[Shot 1] Wide shot. He speaks. (S1) <d>[English] First.</d> He pauses. (S1) <d>[English] Later.</d>',
+    );
+  });
+});
diff --git a/apps/fastvideo_studio/src/components/jobs/H3ScriptImportDialog.test.tsx b/apps/fastvideo_studio/src/components/jobs/H3ScriptImportDialog.test.tsx
new file mode 100644
index 000000000..61d21c3ef
--- /dev/null
+++ b/apps/fastvideo_studio/src/components/jobs/H3ScriptImportDialog.test.tsx
@@ -0,0 +1,123 @@
+import { render, screen, within } from '@testing-library/react';
+import userEvent from '@testing-library/user-event';
+import { describe, expect, it, vi } from 'vitest';
+
+import { H3ShotListEditor } from './H3ShotListEditor';
+
+// Original test material -- not from any real screenplay.
+const SCRIPT = [
+  'CAPTAIN: Sit down. (EXHALES) We need to talk about the ledger.',
+  '',
+  'CLERK: Yes, sir.',
+  '',
+  'CAPTAIN: Three crates are missing and nobody on this dock saw a thing.',
+].join('\n');
+
+const SUBJECTS = '<Subject 1> is the captain.\n<Subject 2> is the clerk.';
+
+function renderEditor(detailed = '', onApply = vi.fn()) {
+  render(
+    <H3ShotListEditor
+      open
+      subjectDefinitions={SUBJECTS}
+      initialDetailedDescription={detailed}
+      initialRetentionAnalysis=""
+      onApply={onApply}
+      onClose={() => {}}
+    />,
+  );
+  return onApply;
+}
+
+async function importScript(user: ReturnType<typeof userEvent.setup>) {
+  await user.click(screen.getByRole('button', { name: 'Import script…' }));
+  const dialog = await screen.findByRole('dialog', { name: 'Import script' });
+  await user.click(within(dialog).getByLabelText('Script'));
+  await user.paste(SCRIPT);
+  return dialog;
+}
+
+describe('script import', () => {
+  it('maps speakers to subjects by default and previews the clip', async () => {
+    const user = userEvent.setup();
+    renderEditor();
+    const dialog = await importScript(user);
+
+    expect(within(dialog).getByLabelText('CAPTAIN subject')).toHaveValue('<Subject 1>');
+    expect(within(dialog).getByLabelText('CLERK subject')).toHaveValue('<Subject 2>');
+    expect(within(dialog).getByLabelText('CAPTAIN speaker tag')).toHaveValue('S1');
+    expect(within(dialog).getByLabelText('CLERK speaker tag')).toHaveValue('S2');
+    expect(within(dialog).getByRole('option', { name: /Clip 1 of \d+/ })).toBeInTheDocument();
+    expect(within(dialog).getByText('Establishing')).toBeInTheDocument();
+  });
+
+  it('loads the chosen clip into the shot list and writes it on apply', async () => {
+    const user = userEvent.setup();
+    const onApply = renderEditor();
+    const dialog = await importScript(user);
+
+    await user.click(within(dialog).getByRole('button', { name: 'Load clip' }));
+
+    const boxes = (await screen.findAllByPlaceholderText('What the speaker says')) as HTMLTextAreaElement[];
+    expect(boxes[0].value).toBe('Sit down. <exhale> We need to talk about the ledger.');
+    expect(boxes[1].value).toBe('Yes, sir.');
+
+    await user.click(screen.getByRole('button', { name: 'Apply' }));
+    const [detailed, retention] = onApply.mock.calls[0];
+    expect(detailed).toContain('[Shot 1] Establishing shot.');
+    expect(detailed).toContain('(S1) <d>[English] Sit down. <exhale> We need to talk about the ledger.</d> (S2) <d>[English] Yes, sir.</d>');
+    expect(retention).toContain('<Subject 1> (appears in');
+    expect(retention).toContain('<Subject 2> (appears in');
+  });
+
+  it('warns that loading replaces the shots already in the list', async () => {
+    const user = userEvent.setup();
+    renderEditor('[Shot 1] Wide shot. A dock.\n\n[Shot 2] Close-up shot. A face.');
+    const dialog = await importScript(user);
+    expect(within(dialog).getByText('Replaces the 2 shots currently in the list.')).toBeInTheDocument();
+  });
+
+  it('reloading the same clip after edits restores the imported text', async () => {
+    const user = userEvent.setup();
+    renderEditor();
+    let dialog = await importScript(user);
+    await user.click(within(dialog).getByRole('button', { name: 'Load clip' }));
+
+    let boxes = (await screen.findAllByPlaceholderText('What the speaker says')) as HTMLTextAreaElement[];
+    await user.type(boxes[1], ' EDITED');
+    expect(boxes[1].value).toBe('Yes, sir. EDITED');
+
+    await user.click(screen.getByRole('button', { name: 'Import script…' }));
+    dialog = await screen.findByRole('dialog', { name: 'Import script' });
+    // Loading a clip advances the picker to the next one; choose clip 1 again.
+    await user.selectOptions(within(dialog).getByLabelText('Clip to load'), '0');
+    await user.click(within(dialog).getByRole('button', { name: 'Load clip' }));
+
+    boxes = (await screen.findAllByPlaceholderText('What the speaker says')) as HTMLTextAreaElement[];
+    expect(boxes[1].value).toBe('Yes, sir.');
+  });
+
+  it('advances to the next clip after loading one', async () => {
+    const user = userEvent.setup();
+    renderEditor();
+    let dialog = await importScript(user);
+    const picker = within(dialog).getByLabelText('Clip to load') as HTMLSelectElement;
+    expect(picker.options.length).toBeGreaterThan(1);
+    await user.click(within(dialog).getByRole('button', { name: 'Load clip' }));
+
+    await user.click(screen.getByRole('button', { name: 'Import script…' }));
+    dialog = await screen.findByRole('dialog', { name: 'Import script' });
+    expect((within(dialog).getByLabelText('Clip to load') as HTMLSelectElement).value).toBe('1');
+  });
+
+  it('explains when the text has no speaker turns', async () => {
+    const user = userEvent.setup();
+    renderEditor();
+    await user.click(screen.getByRole('button', { name: 'Import script…' }));
+    const dialog = await screen.findByRole('dialog', { name: 'Import script' });
+    await user.click(within(dialog).getByLabelText('Script'));
+    await user.paste('just some prose');
+    expect(within(dialog).getByRole('alert')).toHaveTextContent(/No "NAME: line" turns found/);
+    expect(within(dialog).getByRole('button', { name: 'Load clip' })).toBeDisabled();
+  });
+});
diff --git a/apps/fastvideo_studio/src/components/jobs/H3ScriptImportDialog.tsx b/apps/fastvideo_studio/src/components/jobs/H3ScriptImportDialog.tsx
new file mode 100644
index 000000000..a165f6e18
--- /dev/null
+++ b/apps/fastvideo_studio/src/components/jobs/H3ScriptImportDialog.tsx
@@ -0,0 +1,247 @@
+'use client';
+
+import * as React from 'react';
+
+import { Button } from '@/components/ui/button';
+import {
+  Dialog,
+  DialogContent,
+  DialogDescription,
+  DialogFooter,
+  DialogHeader,
+  DialogTitle,
+} from '@/components/ui/dialog';
+import { Input } from '@/components/ui/input';
+import { NativeSelect } from '@/components/ui/native-select';
+import { Textarea } from '@/components/ui/textarea';
+import { buildClips, parseScript, scriptSpeakers, type ImportSpeaker } from '@/lib/h3ScriptImport';
+import { SHOT_TYPES, shotLines, shotProse, type Shot } from '@/lib/h3Shots';
+
+const TYPE_LABEL = Object.fromEntries(SHOT_TYPES.map((t) => [t.id, t.label]));
+const DEFAULT_CLIP_SECONDS = 11;
+
+export interface H3ScriptImportDialogProps {
+  open: boolean;
+  subjectLabels: string[];
+  /** shots currently in the list -- loading a clip replaces them */
+  existingShotCount: number;
+  onLoad: (shots: Shot[]) => void;
+  onClose: () => void;
+}
+
+function preview(shot: Shot): string {
+  const line = shotLines(shot).find((l) => l.text.trim());
+  const text = line ? `${line.speaker ? `${line.speaker}: ` : ''}“${line.text.trim()}”` : shotProse(shot);
+  return text.length > 90 ? `${text.slice(0, 89)}…` : text;
+}
+
+/**
+ * Paste (or load) a "NAME: line" script and pick one clip-sized chunk of it
+ * to load into the shot list. The script and speaker mapping persist between
+ * openings, so working through a long scene is: load clip, apply, next clip.
+ */
+export function H3ScriptImportDialog({
+  open,
+  subjectLabels,
+  existingShotCount,
+  onLoad,
+  onClose,
+}: H3ScriptImportDialogProps) {
+  const [script, setScript] = React.useState('');
+  const [clipSecondsInput, setClipSecondsInput] = React.useState(String(DEFAULT_CLIP_SECONDS));
+  const [overrides, setOverrides] = React.useState<Record<string, Partial<ImportSpeaker>>>({});
+  const [clipIndex, setClipIndex] = React.useState(0);
+
+  const turns = React.useMemo(() => parseScript(script), [script]);
+  const names = React.useMemo(() => scriptSpeakers(turns), [turns]);
+
+  const speakers = React.useMemo(() => {
+    const map: Record<string, ImportSpeaker> = {};
+    names.forEach((name, i) => {
+      const o = overrides[name] ?? {};
+      map[name] = {
+        subject: 'subject' in o ? (o.subject ?? null) : (subjectLabels[i] ?? null),
+        tag: o.tag ?? `S${i + 1}`,
+      };
+    });
+    return map;
+  }, [names, overrides, subjectLabels]);
+
+  const parsedSeconds = Number(clipSecondsInput);
+  const targetSeconds = Number.isFinite(parsedSeconds) ? Math.min(Math.max(parsedSeconds, 4), 12) : DEFAULT_CLIP_SECONDS;
+
+  const clips = React.useMemo(
+    () => (turns.length ? buildClips(turns, { speakers, targetSeconds }) : []),
+    [turns, speakers, targetSeconds],
+  );
+  const selected = Math.min(clipIndex, Math.max(clips.length - 1, 0));
+  const clip = clips[selected];
+
+  function setOverride(name: string, patch: Partial<ImportSpeaker>) {
+    setOverrides((prev) => ({ ...prev, [name]: { ...prev[name], ...patch } }));
+  }
+
+  async function handleFile(e: React.ChangeEvent<HTMLInputElement>) {
+    const file = e.target.files?.[0];
+    e.target.value = '';
+    if (!file) return;
+    setScript(await file.text());
+    setClipIndex(0);
+  }
+
+  function handleLoad() {
+    if (!clip) return;
+    onLoad(clip.shots);
+    setClipIndex(Math.min(selected + 1, clips.length - 1));
+    onClose();
+  }
+
+  return (
+    <Dialog open={open} onOpenChange={(o) => { if (!o) onClose(); }}>
+      <DialogContent className="flex max-h-[90vh] w-[94vw] max-w-[760px] flex-col gap-0 overflow-hidden p-0">
+        <DialogHeader className="border-b border-border px-5 py-4">
+          <DialogTitle>Import script</DialogTitle>
+          <DialogDescription>
+            Paste a script of NAME: line turns. It&apos;s split into clip-sized chunks; load one at a time into the
+            shot list, then set the camera and descriptions.
+          </DialogDescription>
+        </DialogHeader>
+
+        <div className="flex min-h-0 flex-1 flex-col gap-4 overflow-y-auto p-5">
+          <div className="flex flex-col gap-2">
+            <label htmlFor="script-import-text" className="text-xs font-semibold text-foreground">
+              Script
+            </label>
+            <Textarea
+              id="script-import-text"
+              value={script}
+              onChange={(e) => {
+                setScript(e.target.value);
+                setClipIndex(0);
+              }}
+              rows={8}
+              spellCheck={false}
+              placeholder={'MARK: First line...\n\nJORDAN: Reply...'}
+              className="font-mono text-xs"
+            />
+            <Input
+              type="file"
+              accept=".md,.txt,.text"
+              aria-label="Load script file"
+              onChange={handleFile}
+              className="h-auto py-2 file:mr-3 file:cursor-pointer file:rounded-md file:border-0 file:bg-secondary file:px-2 file:py-1 file:text-sm file:text-secondary-foreground"
+            />
+          </div>
+
+          {script.trim() && turns.length === 0 && (
+            <p role="alert" className="text-sm text-destructive">
+              No &quot;NAME: line&quot; turns found. Each turn should start with a speaker name and a colon.
+            </p>
+          )}
+
+          {names.length > 0 && (
+            <div className="flex flex-col gap-2">
+              <span className="text-xs font-semibold text-foreground">Speakers</span>
+              {names.map((name) => (
+                <div key={name} className="grid grid-cols-[1fr_1fr_5rem] items-center gap-2">
+                  <span className="truncate text-sm text-foreground">{name}</span>
+                  <NativeSelect
+                    aria-label={`${name} subject`}
+                    value={speakers[name].subject ?? ''}
+                    onChange={(e) => setOverride(name, { subject: e.target.value || null })}
+                  >
+                    <option value="">Not a tracked subject</option>
+                    {subjectLabels.map((label) => (
+                      <option key={label} value={label}>{label}</option>
+                    ))}
+                  </NativeSelect>
+                  <Input
+                    aria-label={`${name} speaker tag`}
+                    value={speakers[name].tag}
+                    onChange={(e) => setOverride(name, { tag: e.target.value })}
+                    className="font-mono"
+                  />
+                </div>
+              ))}
+              {subjectLabels.length === 0 && (
+                <p className="text-xs text-muted-foreground">
+                  No subjects are defined yet, so speakers can&apos;t be linked. Add them under Subject definitions
+                  to track who appears in each shot.
+                </p>
+              )}
+            </div>
+          )}
+
+          {turns.length > 0 && (
+            <>
+              <label className="flex flex-col gap-1 text-xs text-muted-foreground">
+                Clip length target (seconds)
+                <Input
+                  type="number"
+                  min={4}
+                  max={12}
+                  step={1}
+                  value={clipSecondsInput}
+                  onChange={(e) => {
+                    setClipSecondsInput(e.target.value);
+                    setClipIndex(0);
+                  }}
+                  className="w-28"
+                />
+                <span>One generation runs about 11 s (277 frames at 24 fps), so each clip becomes its own job.</span>
+              </label>
+
+              <div className="flex flex-col gap-2">
+                <label htmlFor="script-import-clip" className="text-xs font-semibold text-foreground">
+                  Clip to load
+                </label>
+                <NativeSelect
+                  id="script-import-clip"
+                  value={selected}
+                  onChange={(e) => setClipIndex(Number(e.target.value))}
+                >
+                  {clips.map((c, i) => (
+                    <option key={i} value={i}>
+                      Clip {i + 1} of {clips.length} · {c.shots.length} shot{c.shots.length === 1 ? '' : 's'}
+                      {c.continuation ? ' · continues previous shot' : ''}
+                    </option>
+                  ))}
+                </NativeSelect>
+
+                {clip?.continuation && (
+                  <p className="rounded-lg border border-border bg-secondary/30 px-3 py-2 text-xs text-muted-foreground">
+                    This clip continues the previous clip&apos;s last shot, cut short because that speech didn&apos;t
+                    fit one generation. Generate it from the previous clip&apos;s last frame (as the start Image)
+                    with the same references.
+                  </p>
+                )}
+
+                {clip && (
+                  <ol className="flex flex-col gap-1 rounded-xl border border-border bg-secondary/20 p-3">
+                    {clip.shots.map((shot, i) => (
+                      <li key={shot.id} className="flex gap-2 text-xs">
+                        <span className="w-6 shrink-0 font-mono text-muted-foreground">{i + 1}</span>
+                        <span className="w-24 shrink-0 font-medium text-foreground">{TYPE_LABEL[shot.type]}</span>
+                        <span className="min-w-0 text-muted-foreground">{preview(shot)}</span>
+                      </li>
+                    ))}
+                  </ol>
+                )}
+              </div>
+            </>
+          )}
+        </div>
+
+        <DialogFooter className="items-center border-t border-border px-5 py-3">
+          {existingShotCount > 0 && clip && (
+            <span className="mr-auto text-xs text-muted-foreground">
+              Replaces the {existingShotCount} shot{existingShotCount === 1 ? '' : 's'} currently in the list.
+            </span>
+          )}
+          <Button type="button" variant="outline" onClick={onClose}>Cancel</Button>
+          <Button type="button" onClick={handleLoad} disabled={!clip}>Load clip</Button>
+        </DialogFooter>
+      </DialogContent>
+    </Dialog>
+  );
+}
diff --git a/apps/fastvideo_studio/src/components/jobs/H3ShotDialogueField.tsx b/apps/fastvideo_studio/src/components/jobs/H3ShotDialogueField.tsx
new file mode 100644
index 000000000..0264f98c6
--- /dev/null
+++ b/apps/fastvideo_studio/src/components/jobs/H3ShotDialogueField.tsx
@@ -0,0 +1,92 @@
+'use client';
+
+import * as React from 'react';
+
+import { Textarea } from '@/components/ui/textarea';
+import { H3DialoguePreview } from '@/components/jobs/H3DialoguePreview';
+import { applyH3Preset, applyH3Tag, type H3Preset, type H3Tag } from '@/lib/h3Tags';
+
+/** What the shared tag palette drives: the last-focused shot dialogue field. */
+export interface H3ShotDialogueHandle {
+  insertTag: (tag: H3Tag) => void;
+  applyPreset: (preset: H3Preset) => void;
+}
+
+export interface H3ShotDialogueFieldProps {
+  /** The line's text. The parent owns it (and the undo history), so this is fully controlled. */
+  value: string;
+  /** Typing. */
+  onChange: (value: string) => void;
+  /** A tag or preset inserted from the palette -- a discrete edit, unlike typing. */
+  onInsert: (value: string) => void;
+  onFocus: () => void;
+  active?: boolean;
+  label?: string;
+}
+
+/**
+ * One spoken line: a textarea plus, once it contains tags, a styled preview.
+ * The shared palette inserts at the cursor through the ref. Undo/redo isn't
+ * here -- the shot list editor keeps one history for everything in it.
+ */
+export const H3ShotDialogueField = React.forwardRef<H3ShotDialogueHandle, H3ShotDialogueFieldProps>(
+  function H3ShotDialogueField({ value, onChange, onInsert, onFocus, active, label = 'Dialogue' }, ref) {
+    const taRef = React.useRef<HTMLTextAreaElement>(null);
+    // Selection to restore once the inserted text has rendered.
+    const pending = React.useRef<[number, number] | null>(null);
+
+    React.useEffect(() => {
+      if (pending.current && taRef.current) {
+        const [s, e] = pending.current;
+        taRef.current.focus();
+        taRef.current.setSelectionRange(s, e);
+        pending.current = null;
+      }
+    });
+
+    function selection(): [number, number] {
+      const ta = taRef.current;
+      return [ta?.selectionStart ?? value.length, ta?.selectionEnd ?? value.length];
+    }
+
+    React.useImperativeHandle(ref, () => ({
+      insertTag(tag) {
+        const [s, e] = selection();
+        const result = applyH3Tag(value, s, e, tag);
+        pending.current = [result.start, result.end];
+        onInsert(result.value);
+      },
+      applyPreset(preset) {
+        const [s, e] = selection();
+        const result = applyH3Preset(value, s, e, preset);
+        pending.current = [result.start, result.end];
+        onInsert(result.value);
+      },
+    }));
+
+    return (
+      <div className="flex flex-col gap-1">
+        <div className="text-xs text-muted-foreground">
+          {label}
+          {active && <span className="ml-2 text-accent-blue">tags insert here</span>}
+        </div>
+        <Textarea
+          ref={taRef}
+          value={value}
+          onChange={(e) => onChange(e.target.value)}
+          onFocus={onFocus}
+          rows={2}
+          spellCheck={false}
+          placeholder="What the speaker says"
+          className="font-mono text-sm"
+        />
+        {value.includes('<') && (
+          <H3DialoguePreview
+            text={value}
+            className="rounded-lg border border-border bg-card px-3 py-2 text-sm leading-7 text-foreground"
+          />
+        )}
+      </div>
+    );
+  },
+);
diff --git a/apps/fastvideo_studio/src/components/jobs/H3ShotListEditor.tsx b/apps/fastvideo_studio/src/components/jobs/H3ShotListEditor.tsx
new file mode 100644
index 000000000..256b3fb9a
--- /dev/null
+++ b/apps/fastvideo_studio/src/components/jobs/H3ShotListEditor.tsx
@@ -0,0 +1,690 @@
+'use client';
+
+import * as React from 'react';
+
+import { Button } from '@/components/ui/button';
+import {
+  Dialog,
+  DialogContent,
+  DialogDescription,
+  DialogFooter,
+  DialogHeader,
+  DialogTitle,
+} from '@/components/ui/dialog';
+import { Input } from '@/components/ui/input';
+import { NativeSelect } from '@/components/ui/native-select';
+import { Textarea } from '@/components/ui/textarea';
+import { H3ShotDialogueField, type H3ShotDialogueHandle } from '@/components/jobs/H3ShotDialogueField';
+import { H3ScriptImportDialog } from '@/components/jobs/H3ScriptImportDialog';
+import { H3ShotStoryboard } from '@/components/jobs/H3ShotStoryboard';
+import { H3TagPalette } from '@/components/jobs/H3TagPalette';
+import type { H3Preset, H3Tag } from '@/lib/h3Tags';
+import {
+  CAMERA_MOVEMENTS,
+  SHOT_TYPES,
+  detailedDescriptionFromShots,
+  formatCutTime,
+  linePart,
+  newLine,
+  newShot,
+  newTextPart,
+  parseCutTime,
+  parseDetailedDescription,
+  parseSubjectLabels,
+  resolveShotStarts,
+  retentionAnalysisFromShots,
+  shotLines,
+  withFreshIds,
+  type CameraMovement,
+  type DialogueLine,
+  type LinePart,
+  type ShotPart,
+  type Shot,
+  type ShotType,
+} from '@/lib/h3Shots';
+import { useUndoableState } from '@/hooks/useUndoableState';
+import { cn } from '@/lib/utils';
+
+export interface H3ShotListEditorProps {
+  open: boolean;
+  /** Raw subject_definitions text -- parsed for <Subject N> labels to track per shot. */
+  subjectDefinitions: string;
+  initialDetailedDescription: string;
+  initialRetentionAnalysis: string;
+  onApply: (detailedDescription: string, retentionAnalysis: string) => void;
+  onClose: () => void;
+  /** Where the shots get written, for the description text. */
+  targetLabel?: string;
+  /** False when there is no retention analysis to keep in sync (base-format and plain prompts). */
+  withRetention?: boolean;
+}
+
+/**
+ * Structured camera-direction editor for MiniMax-H3: an ordered shot list
+ * (type, movement, which subjects appear, action, optional dialogue) that
+ * writes into detailed_description's [Shot N] blocks and keeps
+ * retention_analysis's "appears in [...]" lines in sync automatically,
+ * instead of hand-maintaining both fields' free text in agreement.
+ */
+/** Cut-in time field: blank means "automatic" (shown as the placeholder); Enter or blur commits. */
+function CutTimeInput({
+  value,
+  auto,
+  disabled,
+  onCommit,
+}: {
+  value: number | null;
+  auto: number;
+  disabled?: boolean;
+  onCommit: (seconds: number | null) => void;
+}) {
+  const show = (v: number | null) => (v == null ? '' : formatCutTime(v));
+  const [text, setText] = React.useState(show(value));
+  React.useEffect(() => setText(show(value)), [value]);
+
+  function commit() {
+    const trimmed = text.trim();
+    if (!trimmed) return onCommit(null);
+    const parsed = parseCutTime(trimmed);
+    if (parsed == null) return setText(show(value));
+    onCommit(parsed);
+  }
+
+  return (
+    <Input
+      value={text}
+      onChange={(e) => setText(e.target.value)}
+      onBlur={commit}
+      onKeyDown={(e) => {
+        if (e.key === 'Enter') {
+          e.preventDefault();
+          commit();
+        }
+      }}
+      placeholder={disabled ? 'Opening shot' : formatCutTime(auto)}
+      disabled={disabled}
+      inputMode="decimal"
+      className="font-mono"
+    />
+  );
+}
+
+interface EditorState {
+  shots: Shot[];
+  preamble: string;
+}
+
+export function H3ShotListEditor({
+  open,
+  subjectDefinitions,
+  initialDetailedDescription,
+  initialRetentionAnalysis,
+  onApply,
+  onClose,
+  targetLabel = 'Detailed description',
+  withRetention = true,
+}: H3ShotListEditorProps) {
+  // One history covers everything editable here: the shots (structure, fields and
+  // every line of text) and the scene notes. Typing carries a key so a burst
+  // merges into one step; everything else is a step of its own.
+  const history = useUndoableState<EditorState>({ shots: [], preamble: '' });
+  const { shots, preamble } = history.state;
+  // Returning the same list (or text) hands the history back the same state object,
+  // which it treats as "nothing changed" and does not record.
+  const editShots = (fn: (shots: Shot[]) => Shot[], key: string | null = null) =>
+    history.set((s) => {
+      const next = fn(s.shots);
+      return next === s.shots ? s : { ...s, shots: next };
+    }, key);
+  const setPreamble = (value: string) =>
+    history.set((s) => (s.preamble === value ? s : { ...s, preamble: value }), 'preamble');
+  const [selectedId, setSelectedId] = React.useState<string | null>(null);
+  // The shared tag palette inserts into whichever shot's dialogue was last focused.
+  const [dialogueTargetId, setDialogueTargetId] = React.useState<string | null>(null);
+  const dialogueFields = React.useRef(new Map<string, H3ShotDialogueHandle>());
+  const [importOpen, setImportOpen] = React.useState(false);
+  const subjectLabels = React.useMemo(() => parseSubjectLabels(subjectDefinitions), [subjectDefinitions]);
+
+  // Re-seed from the field's current text each time the editor opens, the
+  // same "seed on the open transition only" guard H3DialogueEditor uses --
+  // otherwise a stray parent re-render while open could clobber in-progress edits.
+  const wasOpenRef = React.useRef(false);
+  React.useEffect(() => {
+    if (open && !wasOpenRef.current) {
+      const { preamble: seededPreamble, shots: seeded } = parseDetailedDescription(
+        initialDetailedDescription,
+        subjectLabels,
+      );
+      history.reset({ preamble: seededPreamble, shots: seeded });
+      setSelectedId(seeded[0]?.id ?? null);
+      setDialogueTargetId(null);
+    }
+    wasOpenRef.current = open;
+    // eslint-disable-next-line react-hooks/exhaustive-deps
+  }, [open]);
+
+  function addShot() {
+    const shot = newShot();
+    editShots((prev) => [...prev, shot]);
+    selectShot(shot.id);
+  }
+
+  // Storyboard panel -> jump to that shot's editor card.
+  function selectShot(id: string) {
+    setSelectedId(id);
+    requestAnimationFrame(() =>
+      document.getElementById(`shot-card-${id}`)?.scrollIntoView({ behavior: 'smooth', block: 'nearest' }),
+    );
+  }
+
+  function removeShot(id: string) {
+    editShots((prev) => prev.filter((s) => s.id !== id));
+    setDialogueTargetId((cur) => (cur === id ? null : cur));
+  }
+
+  function insertTag(tag: H3Tag) {
+    if (dialogueTargetId) dialogueFields.current.get(dialogueTargetId)?.insertTag(tag);
+  }
+
+  function applyPreset(preset: H3Preset) {
+    if (dialogueTargetId) dialogueFields.current.get(dialogueTargetId)?.applyPreset(preset);
+  }
+
+  const partKey = (p: ShotPart) => (p.kind === 'text' ? p.id : p.line.id);
+
+  // An edit that changes nothing returns the same list and so records no undo step
+  // (a blur that re-commits the same value, for instance).
+  function updateBody(shotId: string, fn: (body: ShotPart[]) => ShotPart[], editKey: string | null = null) {
+    editShots((prev) => {
+      let changed = false;
+      const next = prev.map((s) => {
+        if (s.id !== shotId) return s;
+        const body = fn(s.body);
+        if (body === s.body) return s;
+        changed = true;
+        return { ...s, body };
+      });
+      return changed ? next : prev;
+    }, editKey);
+  }
+
+  function addLine(shotId: string) {
+    updateBody(shotId, (body) => {
+      const lastSpeaker = [...body].reverse().find((p): p is LinePart => p.kind === 'line')?.line.speaker;
+      return [...body, linePart(newLine({ speaker: lastSpeaker ?? 'S1' }))];
+    });
+  }
+
+  function addText(shotId: string) {
+    updateBody(shotId, (body) => [...body, newTextPart()]);
+  }
+
+  function removePart(shotId: string, key: string) {
+    updateBody(shotId, (body) => body.filter((p) => partKey(p) !== key));
+    setDialogueTargetId((cur) => (cur === key ? null : cur));
+  }
+
+  function movePart(shotId: string, key: string, dir: -1 | 1) {
+    updateBody(shotId, (body) => {
+      const i = body.findIndex((p) => partKey(p) === key);
+      const j = i + dir;
+      if (i < 0 || j < 0 || j >= body.length) return body;
+      const next = [...body];
+      [next[i], next[j]] = [next[j], next[i]];
+      return next;
+    });
+  }
+
+  function updateText(shotId: string, key: string, text: string) {
+    updateBody(
+      shotId,
+      (body) => {
+        const part = body.find((p) => p.kind === 'text' && p.id === key);
+        if (!part || (part.kind === 'text' && part.text === text)) return body;
+        return body.map((p) => (p === part ? { ...part, text } : p));
+      },
+      `text:${key}`,
+    );
+  }
+
+  /** `editKey` groups typing into one undo step; omit it for a discrete edit like a tag insertion. */
+  function updateLine(shotId: string, lineId: string, patch: Partial<DialogueLine>, editKey?: string) {
+    updateBody(
+      shotId,
+      (body) => {
+        const part = body.find((p) => p.kind === 'line' && p.line.id === lineId);
+        if (!part || part.kind !== 'line') return body;
+        const unchanged = (Object.keys(patch) as (keyof DialogueLine)[]).every((k) => part.line[k] === patch[k]);
+        return unchanged ? body : body.map((p) => (p === part ? { ...part, line: { ...part.line, ...patch } } : p));
+      },
+      editKey ?? null,
+    );
+  }
+
+  const undo = () => void history.undo();
+  const redo = () => void history.redo();
+
+  // Undo/redo for the whole list, from anywhere inside it -- including from a textbox,
+  // where the browser's own undo would only know about that one box. Not while the
+  // script import (nested inside this dialog) is open: its box keeps native undo.
+  function handleKeyDown(e: React.KeyboardEvent) {
+    if (importOpen || e.altKey || !(e.metaKey || e.ctrlKey)) return;
+    const k = e.key.toLowerCase();
+    if (k === 'z') {
+      e.preventDefault();
+      if (e.shiftKey) redo();
+      else undo();
+    } else if (k === 'y' && !e.shiftKey) {
+      e.preventDefault();
+      redo();
+    }
+  }
+
+  // An imported clip replaces the list wholesale, as one undoable step.
+  function loadImportedShots(imported: Shot[]) {
+    const next = withFreshIds(imported);
+    editShots(() => next);
+    setSelectedId(next[0]?.id ?? null);
+    setDialogueTargetId(null);
+  }
+
+  const dialogueTarget = (() => {
+    for (let i = 0; i < shots.length; i += 1) {
+      const lines = shotLines(shots[i]);
+      const li = lines.findIndex((l) => l.id === dialogueTargetId);
+      if (li >= 0) return { shot: i + 1, line: li + 1, multi: lines.length > 1 };
+    }
+    return null;
+  })();
+
+  function moveShot(id: string, dir: -1 | 1) {
+    editShots((prev) => {
+      const i = prev.findIndex((s) => s.id === id);
+      const j = i + dir;
+      if (i < 0 || j < 0 || j >= prev.length) return prev;
+      const next = [...prev];
+      [next[i], next[j]] = [next[j], next[i]];
+      return next;
+    });
+  }
+
+  function updateShot(id: string, patch: Partial<Shot>) {
+    editShots((prev) => {
+      const shot = prev.find((s) => s.id === id);
+      const unchanged = shot && (Object.keys(patch) as (keyof Shot)[]).every((k) => shot[k] === patch[k]);
+      return !shot || unchanged ? prev : prev.map((s) => (s === shot ? { ...s, ...patch } : s));
+    });
+  }
+
+  function toggleSubject(id: string, label: string) {
+    editShots((prev) =>
+      prev.map((s) =>
+        s.id === id
+          ? {
+              ...s,
+              subjectIds: s.subjectIds.includes(label)
+                ? s.subjectIds.filter((l) => l !== label)
+                : [...s.subjectIds, label],
+            }
+          : s,
+      ),
+    );
+  }
+
+  const starts = React.useMemo(() => resolveShotStarts(shots), [shots]);
+  const previewDetailedDescription = React.useMemo(
+    () => detailedDescriptionFromShots(shots, preamble),
+    [shots, preamble],
+  );
+  const previewRetentionAnalysis = React.useMemo(
+    () => (withRetention ? retentionAnalysisFromShots(shots, subjectLabels, initialRetentionAnalysis) : ''),
+    [withRetention, shots, subjectLabels, initialRetentionAnalysis],
+  );
+
+  function handleApply() {
+    onApply(previewDetailedDescription, previewRetentionAnalysis);
+    onClose();
+  }
+
+  return (
+    <Dialog open={open} onOpenChange={(o) => { if (!o) onClose(); }}>
+      <DialogContent
+        onKeyDown={handleKeyDown}
+        className="flex h-[92vh] max-h-[920px] w-[96vw] max-w-[1100px] flex-col gap-0 overflow-hidden p-0"
+      >
+        <DialogHeader className="border-b border-border px-5 py-4">
+          <DialogTitle>Shot list</DialogTitle>
+          <DialogDescription>
+            Build the scene shot by shot. Applying writes [Shot N] blocks into {targetLabel}
+            {withRetention ? ' and keeps Retention analysis\u2019s \u201cappears in\u201d lines in sync' : ''}.
+          </DialogDescription>
+        </DialogHeader>
+
+        <div className="flex shrink-0 flex-col gap-2 border-b border-border bg-secondary/10 px-4 py-3">
+          <div className="flex items-center justify-between">
+            <span className="text-xs font-semibold text-foreground">Storyboard</span>
+            <span className="flex items-center gap-1.5">
+              <Button
+                type="button"
+                variant="outline"
+                size="sm"
+                onClick={undo}
+                disabled={!history.canUndo}
+                title="Undo (Ctrl/Cmd+Z)"
+              >
+                ↶ Undo
+              </Button>
+              <Button
+                type="button"
+                variant="outline"
+                size="sm"
+                onClick={redo}
+                disabled={!history.canRedo}
+                title="Redo (Ctrl/Cmd+Shift+Z)"
+              >
+                ↷ Redo
+              </Button>
+              <Button type="button" variant="outline" size="sm" onClick={() => setImportOpen(true)}>
+                Import script…
+              </Button>
+            </span>
+          </div>
+          <H3ShotStoryboard shots={shots} selectedId={selectedId} onSelect={selectShot} onAdd={addShot} />
+        </div>
+
+        <div className="flex min-h-0 flex-1 flex-col md:flex-row">
+        <div className="flex min-h-0 min-w-0 flex-1 flex-col gap-4 overflow-y-auto p-4">
+          <label className="flex flex-col gap-1 text-xs text-muted-foreground">
+            Scene notes (optional, written before Shot 1 -- style, lighting, anything true of the whole clip)
+            <Textarea
+              value={preamble}
+              onChange={(e) => setPreamble(e.target.value)}
+              rows={2}
+              placeholder="e.g. Live-action, cinematic, shallow depth of field."
+            />
+          </label>
+
+          {shots.length === 0 && (
+            <p className="rounded-xl border border-dashed border-border p-4 text-sm text-muted-foreground">
+              No shots yet.
+            </p>
+          )}
+
+          {shots.map((shot, i) => (
+            <div
+              key={shot.id}
+              id={`shot-card-${shot.id}`}
+              onFocusCapture={() => setSelectedId(shot.id)}
+              className={cn(
+                'flex flex-col gap-3 rounded-xl border bg-secondary/20 p-3',
+                shot.id === selectedId ? 'border-accent-blue' : 'border-border',
+              )}
+            >
+              <div className="flex items-center justify-between">
+                <span className="text-sm font-semibold text-foreground">
+                  Shot {i + 1}
+                  {i > 0 && (
+                    <span className="ml-2 font-mono text-xs font-normal text-muted-foreground">
+                      cuts in at {formatCutTime(starts[i])}
+                    </span>
+                  )}
+                </span>
+                <div className="flex items-center gap-1">
+                  <Button
+                    type="button"
+                    variant="outline"
+                    size="sm"
+                    onClick={() => moveShot(shot.id, -1)}
+                    disabled={i === 0}
+                    title="Move up"
+                  >
+                    ↑
+                  </Button>
+                  <Button
+                    type="button"
+                    variant="outline"
+                    size="sm"
+                    onClick={() => moveShot(shot.id, 1)}
+                    disabled={i === shots.length - 1}
+                    title="Move down"
+                  >
+                    ↓
+                  </Button>
+                  <Button type="button" variant="outline" size="sm" onClick={() => removeShot(shot.id)}>
+                    Remove
+                  </Button>
+                </div>
+              </div>
+
+              <div className="grid grid-cols-1 gap-2 sm:grid-cols-3">
+                <label className="flex flex-col gap-1 text-xs text-muted-foreground">
+                  Shot type
+                  <NativeSelect
+                    value={shot.type}
+                    onChange={(e) => updateShot(shot.id, { type: e.target.value as ShotType, freeLead: false })}
+                  >
+                    {SHOT_TYPES.map((t) => (
+                      <option key={t.id} value={t.id}>{t.label}</option>
+                    ))}
+                  </NativeSelect>
+                </label>
+                <label className="flex flex-col gap-1 text-xs text-muted-foreground">
+                  Camera movement
+                  <NativeSelect
+                    value={shot.movement}
+                    onChange={(e) => updateShot(shot.id, { movement: e.target.value as CameraMovement, freeLead: false })}
+                  >
+                    {CAMERA_MOVEMENTS.map((m) => (
+                      <option key={m.id} value={m.id}>{m.label}</option>
+                    ))}
+                  </NativeSelect>
+                </label>
+                <label className="flex flex-col gap-1 text-xs text-muted-foreground">
+                  Cut time (blank = automatic)
+                  <CutTimeInput
+                    value={shot.startSeconds}
+                    auto={starts[i]}
+                    disabled={i === 0}
+                    onCommit={(startSeconds) => updateShot(shot.id, { startSeconds })}
+                  />
+                </label>
+              </div>
+
+              {shot.freeLead && (
+                <p className="text-xs text-muted-foreground">
+                  This shot keeps its original wording, so type and movement aren&apos;t written. Changing either
+                  adds a generated lead-in in front of the text below.
+                </p>
+              )}
+
+              {subjectLabels.length > 0 && (
+                <div className="flex flex-col gap-1">
+                  <span className="text-xs text-muted-foreground">Subjects in this shot</span>
+                  <div className="flex flex-wrap gap-1.5">
+                    {subjectLabels.map((label) => {
+                      const active = shot.subjectIds.includes(label);
+                      return (
+                        <button
+                          key={label}
+                          type="button"
+                          aria-pressed={active}
+                          onClick={() => toggleSubject(shot.id, label)}
+                          className={cn(
+                            'rounded-full border px-2 py-0.5 font-mono text-[11px] transition-colors',
+                            active
+                              ? 'border-accent-blue bg-accent-blue/15 text-accent-blue'
+                              : 'border-input bg-card text-muted-foreground hover:bg-accent',
+                          )}
+                        >
+                          {label}
+                        </button>
+                      );
+                    })}
+                  </div>
+                </div>
+              )}
+
+              <div className="flex flex-col gap-2">
+                <span className="text-xs text-muted-foreground">
+                  Description and dialogue, in the order they play. Dialogue is wrapped in &lt;d&gt; on apply.
+                </span>
+                {(() => {
+                  let lineNo = 0;
+                  let textNo = 0;
+                  return shot.body.map((part, bi) => {
+                    const key = partKey(part);
+                    const noun = part.kind === 'line' ? `line ${(lineNo += 1)}` : `text ${(textNo += 1)}`;
+                    const controls = (
+                      <div className="mt-6 flex shrink-0 gap-1">
+                        <Button
+                          type="button"
+                          variant="outline"
+                          size="sm"
+                          onClick={() => movePart(shot.id, key, -1)}
+                          disabled={bi === 0}
+                          aria-label={`Move ${noun} up`}
+                          title="Move up"
+                        >
+                          ↑
+                        </Button>
+                        <Button
+                          type="button"
+                          variant="outline"
+                          size="sm"
+                          onClick={() => movePart(shot.id, key, 1)}
+                          disabled={bi === shot.body.length - 1}
+                          aria-label={`Move ${noun} down`}
+                          title="Move down"
+                        >
+                          ↓
+                        </Button>
+                        <Button
+                          type="button"
+                          variant="outline"
+                          size="sm"
+                          onClick={() => removePart(shot.id, key)}
+                          aria-label={`Remove ${noun}`}
+                          title="Remove"
+                        >
+                          ×
+                        </Button>
+                      </div>
+                    );
+
+                    if (part.kind === 'text') {
+                      return (
+                        <div key={key} className="flex items-start gap-2">
+                          <label className="flex min-w-0 flex-1 flex-col gap-1 text-xs text-muted-foreground">
+                            {textNo === 1 ? 'Description (composition, action, environment)' : `Text ${textNo}`}
+                            <Textarea
+                              value={part.text}
+                              onChange={(e) => updateText(shot.id, key, e.target.value)}
+                              rows={2}
+                              placeholder="What's happening in this shot"
+                            />
+                          </label>
+                          {controls}
+                        </div>
+                      );
+                    }
+
+                    const line = part.line;
+                    return (
+                      <div key={key} className="flex items-start gap-2">
+                        <Input
+                          value={line.speaker}
+                          onChange={(e) => updateLine(shot.id, line.id, { speaker: e.target.value }, `speaker:${line.id}`)}
+                          placeholder="S1"
+                          aria-label={`Line ${lineNo} speaker`}
+                          className="mt-7 w-16 shrink-0 font-mono"
+                        />
+                        <div className="min-w-0 flex-1">
+                          <H3ShotDialogueField
+                            ref={(handle) => {
+                              if (handle) dialogueFields.current.set(line.id, handle);
+                              else dialogueFields.current.delete(line.id);
+                            }}
+                            value={line.text}
+                            onChange={(text) => updateLine(shot.id, line.id, { text }, `line:${line.id}`)}
+                            onInsert={(text) => updateLine(shot.id, line.id, { text })}
+                            onFocus={() => setDialogueTargetId(line.id)}
+                            active={line.id === dialogueTargetId}
+                            label={`Line ${lineNo}`}
+                          />
+                        </div>
+                        {controls}
+                      </div>
+                    );
+                  });
+                })()}
+                <div className="flex gap-2">
+                  <Button type="button" variant="outline" size="sm" onClick={() => addText(shot.id)}>
+                    + Add text
+                  </Button>
+                  <Button type="button" variant="outline" size="sm" onClick={() => addLine(shot.id)}>
+                    + Add line
+                  </Button>
+                </div>
+              </div>
+            </div>
+          ))}
+
+          <Button type="button" variant="outline" onClick={addShot} className="self-start">
+            + Add shot
+          </Button>
+
+          <div className="flex flex-col gap-2 rounded-xl border border-border bg-card p-3">
+            <span className="text-xs font-semibold text-foreground">Preview</span>
+            <div className="flex flex-col gap-1">
+              <span className="text-[11px] font-medium uppercase tracking-wide text-muted-foreground">
+                {targetLabel}
+              </span>
+              <pre className="whitespace-pre-wrap rounded-lg bg-secondary/30 p-2 font-mono text-[11px] text-foreground">
+                {previewDetailedDescription || '—'}
+              </pre>
+            </div>
+            {withRetention && (
+            <div className="flex flex-col gap-1">
+              <span className="text-[11px] font-medium uppercase tracking-wide text-muted-foreground">
+                Retention analysis
+              </span>
+              <pre className="whitespace-pre-wrap rounded-lg bg-secondary/30 p-2 font-mono text-[11px] text-foreground">
+                {previewRetentionAnalysis || '—'}
+              </pre>
+            </div>
+            )}
+          </div>
+        </div>
+
+        <aside className="flex max-h-56 shrink-0 flex-col border-t border-border md:max-h-none md:w-64 md:border-l md:border-t-0">
+          <p className="border-b border-border px-3 py-2 text-xs text-muted-foreground">
+            {dialogueTarget
+              ? `Tags insert into Shot ${dialogueTarget.shot}${dialogueTarget.multi ? `, line ${dialogueTarget.line}` : ''} dialogue`
+              : 'Click into a shot\'s dialogue to insert tags'}
+          </p>
+          <H3TagPalette
+            onInsertTag={insertTag}
+            onApplyPreset={applyPreset}
+            disabled={!dialogueTarget}
+            className="overflow-y-auto p-3"
+          />
+        </aside>
+        </div>
+
+        <DialogFooter className="border-t border-border px-5 py-3">
+          <Button type="button" variant="outline" onClick={onClose}>Cancel</Button>
+          <Button type="button" onClick={handleApply}>Apply</Button>
+        </DialogFooter>
+
+        {/* Nested so Radix treats it as part of this dialog (see CreateJobModal). */}
+      <H3ScriptImportDialog
+        open={importOpen}
+        subjectLabels={subjectLabels}
+        existingShotCount={shots.length}
+        onLoad={loadImportedShots}
+        onClose={() => setImportOpen(false)}
+      />
+      </DialogContent>
+    </Dialog>
+  );
+}
diff --git a/apps/fastvideo_studio/src/components/jobs/H3ShotPanel.tsx b/apps/fastvideo_studio/src/components/jobs/H3ShotPanel.tsx
new file mode 100644
index 000000000..bdff55f99
--- /dev/null
+++ b/apps/fastvideo_studio/src/components/jobs/H3ShotPanel.tsx
@@ -0,0 +1,214 @@
+'use client';
+
+import * as React from 'react';
+
+import {
+  CAMERA_MOVEMENTS,
+  SHOT_TYPES,
+  shotLines,
+  shotProse,
+  type CameraMovement,
+  type Shot,
+  type ShotType,
+} from '@/lib/h3Shots';
+
+const TYPE_LABEL = Object.fromEntries(SHOT_TYPES.map((t) => [t.id, t.label])) as Record<ShotType, string>;
+const MOVEMENT_LABEL = Object.fromEntries(CAMERA_MOVEMENTS.map((m) => [m.id, m.label])) as Record<
+  CameraMovement,
+  string
+>;
+
+// Schematic framing per shot type: `scale` sizes the figures and `baseY` is
+// where their shoulders bottom out in the 160x90 frame, so a wide shot reads
+// as small distant figures and a close-up as a head filling the frame.
+const FRAMING: Partial<Record<ShotType, { scale: number; baseY: number }>> = {
+  establishing: { scale: 0.45, baseY: 72 },
+  wide: { scale: 0.8, baseY: 78 },
+  medium: { scale: 1.5, baseY: 92 },
+  'close-up': { scale: 2.6, baseY: 118 },
+};
+
+function figureXs(count: number): number[] {
+  if (count <= 1) return [80];
+  if (count === 2) return [55, 105];
+  return [38, 80, 122];
+}
+
+function Figure({ x, baseY, scale, opacity }: { x: number; baseY: number; scale: number; opacity: number }) {
+  return (
+    <g fill="currentColor" opacity={opacity}>
+      <circle cx={x} cy={baseY - 30 * scale} r={8 * scale} />
+      <ellipse cx={x} cy={baseY - 12 * scale} rx={14 * scale} ry={12 * scale} />
+    </g>
+  );
+}
+
+export function ShotFrame({ shot }: { shot: Shot }) {
+  const count = Math.min(Math.max(shot.subjectIds.length, 1), 3);
+  const empty = shot.subjectIds.length === 0;
+  const opacity = empty ? 0.25 : 0.6;
+  const framing = FRAMING[shot.type];
+
+  let scene: React.ReactNode;
+  if (shot.type === 'extreme-close-up') {
+    scene = (
+      <g fill="currentColor" opacity={opacity}>
+        <circle cx={80} cy={52} r={62} />
+        <ellipse cx={60} cy={42} rx={7} ry={4} className="fill-background" />
+        <ellipse cx={100} cy={42} rx={7} ry={4} className="fill-background" />
+      </g>
+    );
+  } else if (shot.type === 'over-the-shoulder') {
+    scene = (
+      <>
+        <Figure x={22} baseY={135} scale={3} opacity={0.85} />
+        <Figure x={116} baseY={80} scale={1} opacity={opacity} />
+      </>
+    );
+  } else if (shot.type === 'pov') {
+    scene = (
+      <>
+        <Figure x={80} baseY={72} scale={0.9} opacity={opacity} />
+        <path
+          d="M0 0H160V90H0Z M80 45m-52 0a52 40 0 1 0 104 0a52 40 0 1 0 -104 0"
+          fill="currentColor"
+          fillRule="evenodd"
+          opacity={0.3}
+        />
+        <path d="M74 45H86M80 39V51" stroke="currentColor" strokeWidth={1.5} opacity={0.6} />
+      </>
+    );
+  } else if (shot.type === 'insert') {
+    scene = (
+      <g opacity={0.6} fill="none" stroke="currentColor" strokeWidth={2.5}>
+        <rect x={52} y={16} width={56} height={58} rx={8} />
+        <path d="M64 34H96M64 46H96M64 58H84" strokeLinecap="round" />
+      </g>
+    );
+  } else if (framing) {
+    scene = (
+      <>
+        {shot.type === 'establishing' && (
+          <path d="M0 66H160" stroke="currentColor" strokeWidth={1} opacity={0.35} />
+        )}
+        {figureXs(count).map((x, i) => (
+          <Figure
+            key={x}
+            x={shot.type === 'close-up' ? 80 : x}
+            baseY={framing.baseY}
+            scale={framing.scale}
+            opacity={opacity * (i === 0 ? 1 : 0.75)}
+          />
+        ))}
+      </>
+    );
+  }
+
+  return (
+    <svg viewBox="0 0 160 90" className="block h-full w-full text-foreground" aria-hidden>
+      {scene}
+      <MovementOverlay movement={shot.movement} />
+    </svg>
+  );
+}
+
+function MovementOverlay({ movement }: { movement: CameraMovement }) {
+  const g = (children: React.ReactNode) => (
+    <g
+      className="text-accent-blue"
+      stroke="currentColor"
+      fill="none"
+      strokeWidth={2.5}
+      strokeLinecap="round"
+      strokeLinejoin="round"
+    >
+      {children}
+    </g>
+  );
+  switch (movement) {
+    case 'pan':
+      return g(<path d="M28 80H132M28 80l6-4M28 80l6 4M132 80l-6-4M132 80l-6 4" />);
+    case 'tilt':
+      return g(<path d="M146 18V72M146 18l-4 6M146 18l4 6M146 72l-4-6M146 72l4-6" />);
+    case 'dolly':
+      return g(<path d="M10 86L54 56M150 86L106 56M10 86H150" strokeDasharray="4 4" />);
+    case 'zoom':
+      return g(<path d="M16 26V14H28M132 14H144V26M16 64V76H28M132 76H144V64" />);
+    case 'handheld':
+      return g(
+        <rect x={7} y={7} width={146} height={76} rx={4} strokeDasharray="3 4" transform="rotate(-1.6 80 45)" />,
+      );
+    case 'tracking':
+      return g(
+        <>
+          <path d="M22 82H134M134 82l-6-4M134 82l-6 4" />
+          <circle cx={30} cy={82} r={3} fill="currentColor" />
+        </>,
+      );
+    case 'crane':
+      return g(<path d="M24 80L134 20M134 20h-9M134 20l-3 8" />);
+    default:
+      return null;
+  }
+}
+
+function subjectShortLabel(label: string): string {
+  const m = label.match(/(\d+)/);
+  return m ? `Subj ${m[1]}` : label;
+}
+
+export interface H3ShotPanelProps {
+  shot: Shot;
+  /** the number shown in the badge */
+  number: number | string;
+  /** time printed on the frame, e.g. "00:03.500" */
+  timeLabel: string;
+}
+
+/**
+ * The contents of one storyboard panel: badge, shot type and camera move, a
+ * schematic frame, who is in it, what is said, and the shot's prose. Callers
+ * supply the surrounding card (a button in the editor, plain in the scene view).
+ */
+export function H3ShotPanel({ shot, number, timeLabel }: H3ShotPanelProps) {
+  const spoken = shotLines(shot).filter((l) => l.text.trim());
+  return (
+    <>
+      <div className="flex items-center gap-1.5">
+        <span className="flex size-5 shrink-0 items-center justify-center rounded-full bg-secondary text-[11px] font-semibold text-foreground">
+          {number}
+        </span>
+        <span className="truncate text-xs font-medium text-foreground">{TYPE_LABEL[shot.type]}</span>
+        <span className="ml-auto shrink-0 text-[10px] text-muted-foreground">{MOVEMENT_LABEL[shot.movement]}</span>
+      </div>
+
+      <div className="relative aspect-video overflow-hidden rounded-md border border-border bg-secondary/40">
+        <ShotFrame shot={shot} />
+        <span className="absolute bottom-1 left-1 rounded bg-background/80 px-1 font-mono text-[9px] text-muted-foreground">
+          {timeLabel}
+        </span>
+      </div>
+
+      {shot.subjectIds.length > 0 && (
+        <div className="flex flex-wrap gap-1">
+          {shot.subjectIds.map((label) => (
+            <span
+              key={label}
+              className="rounded-full border border-input bg-secondary/50 px-1.5 py-px font-mono text-[10px] text-muted-foreground"
+            >
+              {subjectShortLabel(label)}
+            </span>
+          ))}
+        </div>
+      )}
+
+      {spoken.length > 0 && (
+        <p className="line-clamp-3 rounded-md bg-accent-blue/10 px-1.5 py-1 text-[11px] italic text-foreground">
+          {spoken.map((l) => `${l.speaker.trim() ? `${l.speaker.trim()}: ` : ''}“${l.text.trim()}”`).join(' ')}
+        </p>
+      )}
+
+      <p className="line-clamp-2 text-[11px] text-muted-foreground">{shotProse(shot) || 'No description yet'}</p>
+    </>
+  );
+}
diff --git a/apps/fastvideo_studio/src/components/jobs/H3ShotStoryboard.test.tsx b/apps/fastvideo_studio/src/components/jobs/H3ShotStoryboard.test.tsx
new file mode 100644
index 000000000..e997220c6
--- /dev/null
+++ b/apps/fastvideo_studio/src/components/jobs/H3ShotStoryboard.test.tsx
@@ -0,0 +1,51 @@
+import { render, screen } from '@testing-library/react';
+import userEvent from '@testing-library/user-event';
+import { describe, expect, it, vi } from 'vitest';
+
+import { H3ShotStoryboard } from './H3ShotStoryboard';
+import { CAMERA_MOVEMENTS, SHOT_TYPES, makeShot, newLine, type Shot, type ShotInit } from '@/lib/h3Shots';
+
+function shot(init: ShotInit): Shot {
+  return makeShot(init);
+}
+
+describe('H3ShotStoryboard', () => {
+  it('lays shots out in order with type, movement, subjects, and dialogue', () => {
+    const shots = [
+      shot({ type: 'establishing', description: 'Wide rooftop.', subjectIds: ['<Subject 1>', '<Subject 2>'] }),
+      shot({ type: 'close-up', movement: 'handheld', lines: [newLine({ speaker: 'S2', text: 'I had no choice.' }), newLine({ speaker: 'S1', text: 'Then go.' })], subjectIds: ['<Subject 2>'] }),
+    ];
+    render(<H3ShotStoryboard shots={shots} selectedId={shots[1].id} onSelect={() => {}} onAdd={() => {}} />);
+
+    expect(screen.getByText('Establishing')).toBeInTheDocument();
+    expect(screen.getByText('Close-up')).toBeInTheDocument();
+    expect(screen.getByText('Handheld')).toBeInTheDocument();
+    expect(screen.getByText(/S2: “I had no choice.” S1: “Then go.”/)).toBeInTheDocument();
+    expect(screen.getAllByText('Subj 2')).toHaveLength(2);
+
+    const panels = screen.getAllByRole('button').filter((b) => b.hasAttribute('aria-current') || b.textContent?.match(/^\d/));
+    expect(panels[0]).toHaveTextContent(/^1Establishing/);
+    expect(screen.getByRole('button', { name: /Close-up/ })).toHaveAttribute('aria-current', 'true');
+  });
+
+  it('selects a shot on click and adds via the trailing tile', async () => {
+    const user = userEvent.setup();
+    const onSelect = vi.fn();
+    const onAdd = vi.fn();
+    const shots = [shot({ description: 'First.' }), shot({ description: 'Second.' })];
+    render(<H3ShotStoryboard shots={shots} selectedId={null} onSelect={onSelect} onAdd={onAdd} />);
+
+    await user.click(screen.getByRole('button', { name: /Second\./ }));
+    expect(onSelect).toHaveBeenCalledWith(shots[1].id);
+    await user.click(screen.getByRole('button', { name: /Add shot/ }));
+    expect(onAdd).toHaveBeenCalled();
+  });
+
+  it('renders every shot type and camera movement without crashing', () => {
+    const shots = SHOT_TYPES.flatMap((t) =>
+      CAMERA_MOVEMENTS.map((m) => shot({ type: t.id, movement: m.id, subjectIds: ['<Subject 1>'] })),
+    );
+    const { container } = render(<H3ShotStoryboard shots={shots} selectedId={null} onSelect={() => {}} onAdd={() => {}} />);
+    expect(container.querySelectorAll('svg')).toHaveLength(SHOT_TYPES.length * CAMERA_MOVEMENTS.length);
+  });
+});
diff --git a/apps/fastvideo_studio/src/components/jobs/H3ShotStoryboard.tsx b/apps/fastvideo_studio/src/components/jobs/H3ShotStoryboard.tsx
new file mode 100644
index 000000000..136a7b576
--- /dev/null
+++ b/apps/fastvideo_studio/src/components/jobs/H3ShotStoryboard.tsx
@@ -0,0 +1,61 @@
+'use client';
+
+import * as React from 'react';
+
+import { H3ShotPanel } from '@/components/jobs/H3ShotPanel';
+import { formatCutTime, resolveShotStarts, type Shot } from '@/lib/h3Shots';
+import { cn } from '@/lib/utils';
+
+export interface H3ShotStoryboardProps {
+  shots: Shot[];
+  selectedId: string | null;
+  onSelect: (id: string) => void;
+  onAdd: () => void;
+}
+
+/**
+ * The shot list as a storyboard: one panel per shot, left to right in
+ * playback order. Panels are schematic (no generated imagery exists yet) --
+ * subject size reflects shot type and an overlay marks camera movement --
+ * so the sequence, who's in frame, and where the dialogue falls read at a
+ * glance without scrolling the per-shot editors.
+ */
+export function H3ShotStoryboard({ shots, selectedId, onSelect, onAdd }: H3ShotStoryboardProps) {
+  const starts = React.useMemo(() => resolveShotStarts(shots), [shots]);
+  return (
+    <div className="flex items-stretch gap-2 overflow-x-auto pb-1" aria-label="Storyboard">
+      {shots.map((shot, i) => {
+        const selected = shot.id === selectedId;
+        return (
+          <React.Fragment key={shot.id}>
+            {i > 0 && (
+              <span className="flex shrink-0 items-center text-muted-foreground/60" aria-hidden>
+                ›
+              </span>
+            )}
+            <button
+              type="button"
+              onClick={() => onSelect(shot.id)}
+              aria-current={selected ? 'true' : undefined}
+              className={cn(
+                'flex w-52 shrink-0 flex-col gap-1.5 rounded-xl border bg-card p-2 text-left transition-colors hover:bg-accent/40',
+                selected ? 'border-accent-blue ring-2 ring-accent-blue/50' : 'border-border',
+              )}
+            >
+              <H3ShotPanel shot={shot} number={i + 1} timeLabel={formatCutTime(starts[i])} />
+            </button>
+          </React.Fragment>
+        );
+      })}
+
+      <button
+        type="button"
+        onClick={onAdd}
+        className="flex min-h-32 w-28 shrink-0 flex-col items-center justify-center gap-1 rounded-xl border border-dashed border-border text-xs text-muted-foreground transition-colors hover:bg-accent/40 hover:text-foreground"
+      >
+        <span className="text-lg leading-none">+</span>
+        Add shot
+      </button>
+    </div>
+  );
+}
diff --git a/apps/fastvideo_studio/src/components/jobs/H3TagConsole.tsx b/apps/fastvideo_studio/src/components/jobs/H3TagConsole.tsx
new file mode 100644
index 000000000..3aac0bbd5
--- /dev/null
+++ b/apps/fastvideo_studio/src/components/jobs/H3TagConsole.tsx
@@ -0,0 +1,57 @@
+'use client';
+
+import * as React from 'react';
+
+import { H3_TAG_CATEGORIES, type H3Tag } from '@/lib/h3Tags';
+import { cn } from '@/lib/utils';
+
+export interface H3TagConsoleProps {
+  /** Called with the tag to insert into whichever field is currently focused. */
+  onInsert: (tag: H3Tag) => void;
+  disabled?: boolean;
+  className?: string;
+}
+
+/**
+ * A microexpression "console" for MiniMax-H3 speech tags: click a tag to
+ * insert it at the cursor of whichever prompt field was last focused. Kept
+ * as plain buttons (not a menu) so every tag stays one click away.
+ */
+export function H3TagConsole({ onInsert, disabled, className }: H3TagConsoleProps) {
+  return (
+    <div
+      className={cn(
+        'flex flex-col gap-1.5 rounded-xl border border-input bg-secondary/40 p-2',
+        className,
+      )}
+    >
+      <span className="px-0.5 text-[11px] font-medium uppercase tracking-wide text-muted-foreground">
+        Speech tags — click to insert at the cursor
+      </span>
+      <div className="flex flex-col gap-1.5">
+        {H3_TAG_CATEGORIES.map((category) => (
+          <div key={category.id} className="flex flex-wrap items-center gap-1">
+            <span className="mr-0.5 w-full shrink-0 text-[10px] text-muted-foreground/80 sm:w-auto">
+              {category.label}
+            </span>
+            {category.tags.map((tag) => (
+              <button
+                key={tag.id}
+                type="button"
+                title={`${tag.description} · e.g. "${tag.example}"`}
+                // Keep the target textarea focused (and its selection intact)
+                // so the click lands where the user was typing, not nowhere.
+                onMouseDown={(e) => e.preventDefault()}
+                onClick={() => onInsert(tag)}
+                disabled={disabled}
+                className="rounded-full border border-input bg-card px-2 py-0.5 font-mono text-[11px] text-foreground transition-colors hover:bg-accent disabled:cursor-not-allowed disabled:opacity-50"
+              >
+                {tag.label}
+              </button>
+            ))}
+          </div>
+        ))}
+      </div>
+    </div>
+  );
+}
diff --git a/apps/fastvideo_studio/src/components/jobs/H3TagPalette.tsx b/apps/fastvideo_studio/src/components/jobs/H3TagPalette.tsx
new file mode 100644
index 000000000..962bfe2f1
--- /dev/null
+++ b/apps/fastvideo_studio/src/components/jobs/H3TagPalette.tsx
@@ -0,0 +1,74 @@
+'use client';
+
+import * as React from 'react';
+
+import { H3_PRESETS, H3_TAG_CATEGORIES, type H3Preset, type H3Tag } from '@/lib/h3Tags';
+import { cn } from '@/lib/utils';
+
+// Reuses the app's existing status hues as a purely visual category key --
+// no status meaning implied, just four distinguishable accents that already
+// work in both themes.
+export const CATEGORY_STYLE: Record<string, { chip: string; dot: string }> = {
+  pace: { chip: 'border-blue-400/35 bg-blue-500/10 text-blue-900 dark:text-blue-100', dot: 'bg-blue-500' },
+  delivery: { chip: 'border-amber-400/35 bg-amber-500/10 text-amber-900 dark:text-amber-100', dot: 'bg-amber-500' },
+  vocal: { chip: 'border-emerald-400/35 bg-emerald-500/10 text-emerald-900 dark:text-emerald-100', dot: 'bg-emerald-500' },
+  body: { chip: 'border-rose-400/35 bg-rose-500/10 text-rose-900 dark:text-rose-100', dot: 'bg-rose-500' },
+};
+
+export interface H3TagPaletteProps {
+  onInsertTag: (tag: H3Tag) => void;
+  onApplyPreset: (preset: H3Preset) => void;
+  disabled?: boolean;
+  className?: string;
+}
+
+/** Full speech-tag palette + emotion presets. Buttons keep the target textarea focused so the click lands at its cursor. */
+export function H3TagPalette({ onInsertTag, onApplyPreset, disabled, className }: H3TagPaletteProps) {
+  return (
+    <div className={className}>
+      {H3_TAG_CATEGORIES.map((cat) => {
+        const style = CATEGORY_STYLE[cat.id] ?? CATEGORY_STYLE.body;
+        return (
+          <div key={cat.id} className="mb-4">
+            <div className="mb-1.5 flex items-center gap-2 px-0.5">
+              <span className={cn('size-2 rounded-full', style.dot)} />
+              <span className="text-xs font-semibold text-foreground">{cat.label}</span>
+            </div>
+            <div className="flex flex-col gap-1">
+              {cat.tags.map((tag) => (
+                <button
+                  key={tag.id}
+                  type="button"
+                  title={tag.example}
+                  disabled={disabled}
+                  onMouseDown={(e) => e.preventDefault()}
+                  onClick={() => onInsertTag(tag)}
+                  className="flex items-baseline justify-between gap-2 rounded-lg border border-input bg-card px-2.5 py-1.5 text-left transition-colors hover:bg-accent disabled:cursor-not-allowed disabled:opacity-50"
+                >
+                  <span className="font-mono text-[11.5px] text-foreground">{tag.label}</span>
+                  <span className="truncate text-[10.5px] text-muted-foreground">{tag.description}</span>
+                </button>
+              ))}
+            </div>
+          </div>
+        );
+      })}
+
+      <div className="mb-1.5 mt-5 px-0.5 text-xs font-semibold text-foreground">Emotion presets</div>
+      <div className="flex flex-wrap gap-1.5">
+        {H3_PRESETS.map((p) => (
+          <button
+            key={p.label}
+            type="button"
+            disabled={disabled}
+            onMouseDown={(e) => e.preventDefault()}
+            onClick={() => onApplyPreset(p)}
+            className="rounded-full border border-input bg-card px-2.5 py-1 text-[11.5px] text-foreground transition-colors hover:bg-accent disabled:cursor-not-allowed disabled:opacity-50"
+          >
+            {p.label}
+          </button>
+        ))}
+      </div>
+    </div>
+  );
+}
diff --git a/apps/fastvideo_studio/src/components/jobs/nestedDialogs.test.tsx b/apps/fastvideo_studio/src/components/jobs/nestedDialogs.test.tsx
new file mode 100644
index 000000000..8dc968172
--- /dev/null
+++ b/apps/fastvideo_studio/src/components/jobs/nestedDialogs.test.tsx
@@ -0,0 +1,158 @@
+import * as React from 'react';
+import { flushSync } from 'react-dom';
+import { render, screen, waitFor, within } from '@testing-library/react';
+import userEvent from '@testing-library/user-event';
+import { beforeEach, describe, expect, it, vi } from 'vitest';
+
+import CreateJobModal from './CreateJobModal';
+import { createJob, getDatasets, getModels } from '@/lib/api';
+import { defaultOptionsStore } from '@/stores/defaultOptions';
+import { DEFAULT_OPTIONS } from '@/lib/defaultOptions';
+
+vi.mock('sonner', () => ({ toast: { warning: vi.fn(), error: vi.fn(), success: vi.fn() } }));
+vi.mock('@/lib/api', () => ({
+  createJob: vi.fn(),
+  getModels: vi.fn(),
+  getDatasets: vi.fn(),
+  uploadImage: vi.fn(),
+  getSettings: vi.fn(),
+  updateSettings: vi.fn(),
+}));
+
+const H3 = { id: 'minimax-h3/i2v', label: 'MiniMax H3', type: 'i2v' };
+
+// A browser runs a microtask checkpoint -- so React flushes -- between the
+// listeners of a single native event; jsdom's synchronous dispatch never does.
+// Radix decides whether a parent dialog should dismiss from the layer stack at
+// the moment its listener runs, so without this the bug can't show up here.
+const wrapped = new WeakMap<object, EventListener>();
+beforeEach(() => {
+  const realAdd = Document.prototype.addEventListener;
+  const realRemove = Document.prototype.removeEventListener;
+  vi.spyOn(document, 'addEventListener').mockImplementation(function (
+    this: Document,
+    type: string,
+    listener: unknown,
+    options?: unknown,
+  ) {
+    if (typeof listener === 'function' && ['keydown', 'pointerdown', 'click'].includes(type)) {
+      const flushing: EventListener = (e) => {
+        (listener as EventListener)(e);
+        flushSync(() => {});
+      };
+      wrapped.set(listener, flushing);
+      return realAdd.call(this, type, flushing, options as AddEventListenerOptions);
+    }
+    return realAdd.call(this, type, listener as EventListener, options as AddEventListenerOptions);
+  } as typeof document.addEventListener);
+  vi.spyOn(document, 'removeEventListener').mockImplementation(function (
+    this: Document,
+    type: string,
+    listener: unknown,
+    options?: unknown,
+  ) {
+    return realRemove.call(
+      this,
+      type,
+      (wrapped.get(listener as object) ?? listener) as EventListener,
+      options as EventListenerOptions,
+    );
+  } as typeof document.removeEventListener);
+
+  defaultOptionsStore.set({ options: DEFAULT_OPTIONS });
+  vi.mocked(getModels).mockResolvedValue([H3]);
+  vi.mocked(getDatasets).mockResolvedValue([]);
+  vi.mocked(createJob).mockResolvedValue({ id: 'j' } as never);
+});
+
+function Harness() {
+  const [open, setOpen] = React.useState(true);
+  return (
+    <>
+      <div data-testid="state">{open ? 'open' : 'closed'}</div>
+      <CreateJobModal
+        isOpen={open}
+        onClose={() => setOpen(false)}
+        onSuccess={() => {}}
+        jobType="inference"
+        workloadType="i2v"
+      />
+    </>
+  );
+}
+
+type User = ReturnType<typeof userEvent.setup>;
+type Closer = (user: User, dialog: HTMLElement) => Promise<void>;
+
+async function renderH3Modal(user: User) {
+  render(<Harness />);
+  await screen.findByRole('option', { name: /MiniMax H3/ });
+  await user.selectOptions(screen.getByLabelText('Model'), H3.id);
+}
+
+// Radix marks every dialog under the topmost one aria-hidden, which also hides its
+// accessible name, so the job dialog is identified by its title text instead.
+const expectParentStillOpen = () => {
+  expect(screen.getByTestId('state')).toHaveTextContent('open');
+  expect(screen.getByText(/New Inference Job/)).toBeInTheDocument();
+};
+// [0] is the job dialog's overlay; the last is the topmost dialog's.
+const topOverlay = () => [...document.querySelectorAll('.bg-black\\/70')].at(-1) as Element;
+
+const CLOSERS: [string, Closer][] = [
+  ['Cancel', (u, d) => u.click(within(d).getByRole('button', { name: 'Cancel' }))],
+  ['the X button', (u, d) => u.click(within(d).getByRole('button', { name: 'Close' }))],
+  ['Escape', (u) => u.keyboard('{Escape}')],
+  ['a click on the overlay', (u) => u.click(topOverlay())],
+];
+
+describe('closing a dialog opened from the job form', () => {
+  describe.each([
+    ['shot list', 'Open shot list', 'Shot list', [...CLOSERS, ['Apply', (u: User, d: HTMLElement) => u.click(within(d).getByRole('button', { name: 'Apply' }))] as [string, Closer]]],
+    ['dialogue editor', 'Open dialogue editor', 'Dialogue editor', [...CLOSERS, ['Apply to field', (u: User, d: HTMLElement) => u.click(within(d).getByRole('button', { name: 'Apply to field' }))] as [string, Closer]]],
+  ] as const)('%s', (_label, openButton, title, closers) => {
+    it.each(closers)('closing with %s leaves the job form open', async (_name, close) => {
+      const user = userEvent.setup();
+      await renderH3Modal(user);
+      await user.click(await screen.findByRole('button', { name: openButton }));
+      const child = await screen.findByRole('dialog', { name: title });
+
+      await close(user, child);
+
+      await waitFor(() => expect(screen.queryByRole('dialog', { name: title })).not.toBeInTheDocument());
+      expectParentStillOpen();
+    });
+  });
+
+  it('the job form is still usable afterwards', async () => {
+    const user = userEvent.setup();
+    await renderH3Modal(user);
+    await user.click(await screen.findByRole('button', { name: 'Open shot list' }));
+    const child = await screen.findByRole('dialog', { name: 'Shot list' });
+    await user.click(within(child).getByRole('button', { name: 'Cancel' }));
+    await waitFor(() => expect(screen.queryByRole('dialog', { name: 'Shot list' })).not.toBeInTheDocument());
+
+    await user.type(screen.getByLabelText('Name (optional)'), 'still here');
+    expect(screen.getByLabelText('Name (optional)')).toHaveValue('still here');
+    // ...and the dialog can be opened again
+    await user.click(screen.getByRole('button', { name: 'Open shot list' }));
+    expect(await screen.findByRole('dialog', { name: 'Shot list' })).toBeInTheDocument();
+  });
+
+  describe('the script import inside the shot list', () => {
+    it.each(CLOSERS.slice(0, 4))('closing it with %s leaves the shot list and the job form open', async (_name, close) => {
+      const user = userEvent.setup();
+      await renderH3Modal(user);
+      await user.click(await screen.findByRole('button', { name: 'Open shot list' }));
+      const shotList = await screen.findByRole('dialog', { name: 'Shot list' });
+      await user.click(within(shotList).getByRole('button', { name: 'Import script…' }));
+      const importDialog = await screen.findByRole('dialog', { name: 'Import script' });
+
+      await close(user, importDialog);
+
+      await waitFor(() => expect(screen.queryByRole('dialog', { name: 'Import script' })).not.toBeInTheDocument());
+      expect(screen.getByRole('dialog', { name: 'Shot list' })).toBeInTheDocument();
+      expectParentStillOpen();
+    });
+  });
+});
diff --git a/apps/fastvideo_studio/src/components/jobs/shotListUndo.test.tsx b/apps/fastvideo_studio/src/components/jobs/shotListUndo.test.tsx
new file mode 100644
index 000000000..b2302df7c
--- /dev/null
+++ b/apps/fastvideo_studio/src/components/jobs/shotListUndo.test.tsx
@@ -0,0 +1,253 @@
+import { render, screen, within } from '@testing-library/react';
+import userEvent from '@testing-library/user-event';
+import { describe, expect, it, vi } from 'vitest';
+
+import { H3ShotListEditor } from './H3ShotListEditor';
+
+const TWO_SHOTS = '[Shot 1] Wide shot. A dock.\n[Shot 2] At 00:02.000, the shot cuts to a close-up shot. A face.';
+const WITH_LINE = '[Shot 1] Wide shot. (S1) <d>[English] Hi.</d>';
+
+type User = ReturnType<typeof userEvent.setup>;
+
+function setup(detailed: string, extra: { open?: boolean } = {}) {
+  const onApply = vi.fn();
+  const ui = (open: boolean) => (
+    <H3ShotListEditor
+      open={open}
+      subjectDefinitions="<Subject 1> is a man."
+      initialDetailedDescription={detailed}
+      initialRetentionAnalysis=""
+      onApply={onApply}
+      onClose={() => {}}
+    />
+  );
+  const view = render(ui(extra.open ?? true));
+  return { onApply, view, ui };
+}
+
+const undoButton = () => screen.getByRole('button', { name: /Undo/ });
+const redoButton = () => screen.getByRole('button', { name: /Redo/ });
+const applied = (onApply: ReturnType<typeof vi.fn>) => onApply.mock.calls.at(-1)?.[0] as string;
+const apply = (user: User) => user.click(screen.getByRole('button', { name: 'Apply' }));
+const lineBox = () => screen.getByPlaceholderText('What the speaker says') as HTMLTextAreaElement;
+const ctrl = (key: string) => `{Control>}${key}{/Control}`;
+
+describe('shot list undo/redo', () => {
+  it('starts with nothing to undo or redo', () => {
+    setup(TWO_SHOTS);
+    expect(undoButton()).toBeDisabled();
+    expect(redoButton()).toBeDisabled();
+  });
+
+  it('undoes and redoes moving a shot', async () => {
+    const user = userEvent.setup();
+    const { onApply } = setup(TWO_SHOTS);
+
+    await user.click(screen.getAllByRole('button', { name: '↑' })[1]);
+    expect(undoButton()).toBeEnabled();
+
+    await user.click(undoButton());
+    expect(redoButton()).toBeEnabled();
+    await apply(user);
+    expect(applied(onApply)).toBe(TWO_SHOTS);
+  });
+
+  it('redo reapplies what was undone', async () => {
+    const user = userEvent.setup();
+    const { onApply } = setup(TWO_SHOTS);
+    await user.click(screen.getAllByRole('button', { name: '↑' })[1]);
+    await user.click(undoButton());
+    await user.click(redoButton());
+    await apply(user);
+    expect(applied(onApply).startsWith('[Shot 1] Close-up shot. A face.')).toBe(true);
+  });
+
+  it('undoes removing a shot, dialogue and all', async () => {
+    const user = userEvent.setup();
+    const { onApply } = setup(`${WITH_LINE}\n[Shot 2] At 00:02.500, the shot cuts to a wide shot. B.`);
+    await user.click(screen.getAllByRole('button', { name: 'Remove' })[0]);
+    expect(screen.queryByPlaceholderText('What the speaker says')).not.toBeInTheDocument();
+
+    await user.click(undoButton());
+    expect(lineBox().value).toBe('Hi.');
+    await apply(user);
+    expect(applied(onApply)).toContain('(S1) <d>[English] Hi.</d>');
+  });
+
+  it('undoes an add, a type change, a subject toggle, and a cut time in turn', async () => {
+    const user = userEvent.setup();
+    const { onApply } = setup(TWO_SHOTS);
+
+    await user.click(screen.getAllByRole('button', { name: /Add shot/ })[0]);
+    await user.selectOptions(screen.getAllByLabelText('Shot type')[0], 'pov');
+    await user.click(screen.getAllByRole('button', { name: '<Subject 1>' })[0]);
+    const cut = screen.getAllByPlaceholderText(/\d\d:\d\d\.\d{3}/)[0] as HTMLInputElement;
+    await user.type(cut, '5{Enter}');
+
+    // four discrete steps, undone one at a time
+    for (let i = 0; i < 4; i += 1) await user.click(undoButton());
+    expect(undoButton()).toBeDisabled();
+    await apply(user);
+    expect(applied(onApply)).toBe(TWO_SHOTS);
+  });
+
+  it('committing a cut time that has not changed is not a step', async () => {
+    const user = userEvent.setup();
+    setup(TWO_SHOTS);
+    const cut = screen.getAllByPlaceholderText(/\d\d:\d\d\.\d{3}/)[0] as HTMLInputElement;
+    await user.type(cut, '5{Enter}');
+    await user.click(undoButton()); // the click blurs the box, which re-commits 00:05.000
+    expect(undoButton()).toBeDisabled();
+  });
+
+  it('a burst of typing is one step', async () => {
+    const user = userEvent.setup();
+    const { onApply } = setup(WITH_LINE);
+    await user.type(lineBox(), ' there my friend');
+    expect(lineBox().value).toBe('Hi. there my friend');
+
+    await user.click(undoButton());
+    expect(lineBox().value).toBe('Hi.');
+    expect(undoButton()).toBeDisabled();
+    await apply(user);
+    expect(applied(onApply)).toContain('<d>[English] Hi.</d>');
+  });
+
+  it('a tag inserted from the palette is its own step, separate from typing', async () => {
+    const user = userEvent.setup();
+    setup(WITH_LINE);
+    await user.type(lineBox(), ' there');
+    lineBox().setSelectionRange(lineBox().value.length, lineBox().value.length);
+    await user.click(screen.getByRole('button', { name: /<pause>/ }));
+    expect(lineBox().value).toBe('Hi. there <pause>');
+
+    await user.click(undoButton());
+    expect(lineBox().value).toBe('Hi. there');
+    await user.click(undoButton());
+    expect(lineBox().value).toBe('Hi.');
+  });
+
+  it('undoes edits to the description and the scene notes', async () => {
+    const user = userEvent.setup();
+    const { onApply } = setup('[Shot 1] Wide shot. A dock.');
+    await user.type(screen.getByPlaceholderText(/Live-action, cinematic/), 'Noir.');
+    await user.type(screen.getByPlaceholderText("What's happening in this shot"), ' At dawn.');
+
+    await user.click(undoButton());
+    await user.click(undoButton());
+    expect(undoButton()).toBeDisabled();
+    await apply(user);
+    expect(applied(onApply)).toBe('[Shot 1] Wide shot. A dock.');
+  });
+
+  it('typing in different boxes makes separate steps', async () => {
+    const user = userEvent.setup();
+    setup('[Shot 1] Wide shot. A dock.');
+    await user.type(screen.getByPlaceholderText(/Live-action, cinematic/), 'Noir.');
+    await user.type(screen.getByPlaceholderText("What's happening in this shot"), ' At dawn.');
+
+    await user.click(undoButton());
+    expect((screen.getByPlaceholderText("What's happening in this shot") as HTMLTextAreaElement).value).toBe('A dock.');
+    expect((screen.getByPlaceholderText(/Live-action, cinematic/) as HTMLTextAreaElement).value).toBe('Noir.');
+  });
+
+  it('a new edit clears what could be redone', async () => {
+    const user = userEvent.setup();
+    setup(TWO_SHOTS);
+    await user.click(screen.getAllByRole('button', { name: '↑' })[1]);
+    await user.click(undoButton());
+    expect(redoButton()).toBeEnabled();
+    await user.click(screen.getAllByRole('button', { name: /Add shot/ })[0]);
+    expect(redoButton()).toBeDisabled();
+  });
+});
+
+describe('shot list undo/redo shortcuts', () => {
+  it('Ctrl+Z from inside a textbox undoes the whole list, not just that box', async () => {
+    const user = userEvent.setup();
+    const { onApply } = setup(TWO_SHOTS);
+    await user.click(screen.getAllByRole('button', { name: '↑' })[1]); // move a shot
+    // focus a different textbox, then undo from there
+    const box = screen.getAllByPlaceholderText("What's happening in this shot")[0] as HTMLTextAreaElement;
+    await user.click(box);
+    await user.keyboard(ctrl('z'));
+    expect(undoButton()).toBeDisabled();
+    await apply(user);
+    expect(applied(onApply)).toBe(TWO_SHOTS);
+  });
+
+  it('redoes with Ctrl+Shift+Z and with Ctrl+Y', async () => {
+    const user = userEvent.setup();
+    setup(TWO_SHOTS);
+    await user.click(screen.getAllByRole('button', { name: '↑' })[1]);
+    const box = screen.getAllByPlaceholderText("What's happening in this shot")[0];
+    await user.click(box);
+
+    await user.keyboard(ctrl('z'));
+    expect(redoButton()).toBeEnabled();
+    await user.keyboard('{Control>}{Shift>}z{/Shift}{/Control}');
+    expect(redoButton()).toBeDisabled();
+
+    await user.keyboard(ctrl('z'));
+    await user.keyboard(ctrl('y'));
+    expect(redoButton()).toBeDisabled();
+    expect(undoButton()).toBeEnabled();
+  });
+
+  it('Cmd+Z works too', async () => {
+    const user = userEvent.setup();
+    setup(TWO_SHOTS);
+    await user.click(screen.getAllByRole('button', { name: '↑' })[1]);
+    await user.click(screen.getAllByPlaceholderText("What's happening in this shot")[0]);
+    await user.keyboard('{Meta>}z{/Meta}');
+    expect(undoButton()).toBeDisabled();
+  });
+
+  it('leaves the script import box to its own undo', async () => {
+    const user = userEvent.setup();
+    const { onApply } = setup(TWO_SHOTS);
+    await user.click(screen.getAllByRole('button', { name: '↑' })[1]);
+    expect(undoButton()).toBeEnabled();
+
+    await user.click(screen.getByRole('button', { name: 'Import script…' }));
+    const dialog = await screen.findByRole('dialog', { name: 'Import script' });
+    const box = within(dialog).getByLabelText('Script');
+    await user.click(box);
+    await user.keyboard(ctrl('z'));
+    await user.click(within(dialog).getByRole('button', { name: 'Cancel' }));
+
+    // the shot list's history was untouched: the move is still there to undo
+    await user.click(await screen.findByRole('button', { name: /Undo/ }));
+    await apply(user);
+    expect(applied(onApply)).toBe(TWO_SHOTS);
+  });
+});
+
+describe('shot list history across imports and reopening', () => {
+  it('loading a script clip is one undoable step', async () => {
+    const user = userEvent.setup();
+    const { onApply } = setup(TWO_SHOTS);
+    await user.click(screen.getByRole('button', { name: 'Import script…' }));
+    const dialog = await screen.findByRole('dialog', { name: 'Import script' });
+    await user.click(within(dialog).getByLabelText('Script'));
+    await user.paste('CAPTAIN: Sit down and hear me out.\n\nCLERK: Yes, sir.');
+    await user.click(within(dialog).getByRole('button', { name: 'Load clip' }));
+    await screen.findAllByPlaceholderText('What the speaker says');
+
+    await user.click(undoButton());
+    expect(screen.queryByPlaceholderText('What the speaker says')).not.toBeInTheDocument();
+    await apply(user);
+    expect(applied(onApply)).toBe(TWO_SHOTS);
+  });
+
+  it('opening the editor again starts a fresh history', async () => {
+    const user = userEvent.setup();
+    const { view, ui } = setup(TWO_SHOTS);
+    await user.click(screen.getAllByRole('button', { name: '↑' })[1]);
+    expect(undoButton()).toBeEnabled();
+
+    view.rerender(ui(false));
+    view.rerender(ui(true));
+    expect(await screen.findByRole('button', { name: /Undo/ })).toBeDisabled();
+  });
+});
diff --git a/apps/fastvideo_studio/src/components/scenes/SceneBoard.tsx b/apps/fastvideo_studio/src/components/scenes/SceneBoard.tsx
new file mode 100644
index 000000000..289307926
--- /dev/null
+++ b/apps/fastvideo_studio/src/components/scenes/SceneBoard.tsx
@@ -0,0 +1,249 @@
+'use client';
+
+import * as React from 'react';
+
+import { H3ShotPanel } from '@/components/jobs/H3ShotPanel';
+import { Badge, type BadgeProps } from '@/components/ui/badge';
+import { Button } from '@/components/ui/button';
+import { NativeSelect } from '@/components/ui/native-select';
+import { getJobVideoUrl } from '@/lib/api';
+import { formatCutTime } from '@/lib/h3Shots';
+import { formatClock, type Scene, type SceneClip } from '@/lib/scenes';
+import type { Job } from '@/lib/types';
+import { cn } from '@/lib/utils';
+
+const BADGE_VARIANTS: Record<string, BadgeProps['variant']> = {
+  pending: 'secondary',
+  running: 'warning',
+  completed: 'success',
+  failed: 'destructive',
+  stopped: 'secondary',
+};
+
+const SEGMENT_TONE: Record<string, string> = {
+  completed: 'bg-emerald-500/70',
+  running: 'bg-amber-500/70',
+  failed: 'bg-rose-500/70',
+};
+
+/** Statuses a job can still be edited in (mirrors the Jobs page). */
+export const EDITABLE_STATUSES = ['pending', 'failed', 'stopped'];
+
+function clipLabel(clip: SceneClip): string {
+  return clip.job.name?.trim() || `Job ${clip.job.id.slice(0, 8)}`;
+}
+
+function takeLabel(job: Job, index: number, total: number): string {
+  const ms = job.created_at < 1e12 ? job.created_at * 1000 : job.created_at;
+  const when = new Date(ms).toLocaleString(undefined, {
+    month: 'short',
+    day: 'numeric',
+    hour: '2-digit',
+    minute: '2-digit',
+  });
+  return `Take ${index + 1} of ${total} · ${job.status} · ${when}`;
+}
+
+function scrollToClip(id: string) {
+  document.getElementById(`clip-${id}`)?.scrollIntoView({ behavior: 'smooth', block: 'start' });
+}
+
+function Timeline({ scene }: { scene: Scene }) {
+  return (
+    <div className="flex h-4 w-full gap-px overflow-hidden rounded-md" role="group" aria-label="Scene timeline">
+      {scene.clips.map((clip) => (
+        <button
+          key={clip.job.id}
+          type="button"
+          onClick={() => scrollToClip(clip.job.id)}
+          aria-label={`Go to clip ${clip.position}: ${clipLabel(clip)}`}
+          title={`Clip ${clip.position} · ${clipLabel(clip)} · ${clip.job.status} · ${formatClock(clip.startSeconds)}–${formatClock(clip.startSeconds + clip.durationSeconds)}`}
+          style={{ flexGrow: Math.max(clip.durationSeconds, 0.1), flexBasis: 0 }}
+          className={cn(
+            'min-w-[3px] transition-opacity hover:opacity-70',
+            SEGMENT_TONE[clip.job.status] ?? 'bg-muted-foreground/40',
+          )}
+        />
+      ))}
+    </div>
+  );
+}
+
+function ClipPreview({ job }: { job: Job }) {
+  const isImage = job.output_path?.toLowerCase().endsWith('.png') ?? false;
+  const src = getJobVideoUrl(job.id);
+  return isImage ? (
+    // eslint-disable-next-line @next/next/no-img-element
+    <img src={src} alt={job.name ?? 'Generated output'} className="max-h-56 rounded-md border border-border" />
+  ) : (
+    <video
+      src={src}
+      aria-label={`Generated video for ${job.name ?? job.id}`}
+      className="max-h-56 rounded-md border border-border bg-black"
+      controls
+      muted
+      loop
+      playsInline
+      preload="metadata"
+    />
+  );
+}
+
+function ClipRow({
+  clip,
+  firstShotNumber,
+  onOpenJob,
+  onChooseTake,
+}: {
+  clip: SceneClip;
+  firstShotNumber: number;
+  onOpenJob: (job: Job) => void;
+  onChooseTake: (takeKey: string, jobId: string) => void;
+}) {
+  const [previewing, setPreviewing] = React.useState(false);
+  const { job } = clip;
+  const canPreview = job.status === 'completed' && !!job.output_path;
+  const editable = EDITABLE_STATUSES.includes(job.status);
+
+  return (
+    <li id={`clip-${job.id}`} className="scroll-mt-32 rounded-xl border border-border bg-card p-3">
+      <div className="mb-2 flex flex-wrap items-center gap-2">
+        <span className="flex size-6 shrink-0 items-center justify-center rounded-md bg-secondary text-xs font-semibold text-foreground">
+          {clip.position}
+        </span>
+        <h3 className="min-w-0 truncate text-sm font-semibold text-foreground">{clipLabel(clip)}</h3>
+        <Badge variant={BADGE_VARIANTS[job.status] ?? 'secondary'}>{job.status}</Badge>
+        <span className="font-mono text-xs text-muted-foreground">
+          {formatClock(clip.startSeconds)}–{formatClock(clip.startSeconds + clip.durationSeconds)}
+        </span>
+        {clip.continuation && (
+          <Badge variant="outline" className="normal-case tracking-normal">
+            continues previous shot
+          </Badge>
+        )}
+        <span className="ml-auto flex items-center gap-2">
+          {clip.takeKey && clip.takes.length > 1 && (
+            <NativeSelect
+              aria-label={`Take for clip ${clip.position}`}
+              value={job.id}
+              onChange={(e) => onChooseTake(clip.takeKey as string, e.target.value)}
+              className="h-8 w-auto py-0 text-xs"
+            >
+              {clip.takes.map((take, i) => (
+                <option key={take.id} value={take.id}>
+                  {takeLabel(take, i, clip.takes.length)}
+                </option>
+              ))}
+            </NativeSelect>
+          )}
+          {canPreview && (
+            <Button type="button" size="sm" variant="outline" onClick={() => setPreviewing((p) => !p)}>
+              {previewing ? 'Hide video' : 'Preview'}
+            </Button>
+          )}
+          <Button type="button" size="sm" variant="outline" onClick={() => onOpenJob(job)}>
+            {editable ? 'Edit' : 'View'}
+          </Button>
+        </span>
+      </div>
+
+      {previewing && canPreview && (
+        <div className="mb-2">
+          <ClipPreview job={job} />
+        </div>
+      )}
+
+      {clip.shots.length === 0 ? (
+        <p className="rounded-lg border border-dashed border-border p-3 text-sm text-muted-foreground">
+          This job&apos;s prompt has no shots.
+        </p>
+      ) : (
+        <div className="flex items-stretch gap-2 overflow-x-auto pb-1" aria-label={`Shots in ${clipLabel(clip)}`}>
+          {clip.shots.map((shot, i) => (
+            <div
+              key={shot.id}
+              className="flex w-52 shrink-0 flex-col gap-1.5 rounded-xl border border-border bg-background p-2"
+            >
+              <H3ShotPanel
+                shot={shot}
+                number={firstShotNumber + i}
+                timeLabel={formatCutTime(clip.shotStarts[i])}
+              />
+            </div>
+          ))}
+        </div>
+      )}
+    </li>
+  );
+}
+
+export interface SceneBoardProps {
+  scene: Scene;
+  onOpenJob: (job: Job) => void;
+  /** pick which re-run of a clip number to show */
+  onChooseTake: (takeKey: string, jobId: string) => void;
+}
+
+/** One scene: a timeline of its clips, then every clip's shots in order, numbered straight through. */
+export function SceneBoard({ scene, onOpenJob, onChooseTake }: SceneBoardProps) {
+  let shotNumber = 1;
+  const statuses = Object.entries(scene.statusCounts);
+
+  // The header stays where it is until it reaches the top of the scrolling area,
+  // then sticks there. A 1px sentinel sitting just above it tells us which: once
+  // the sentinel has scrolled out of view, the header is stuck.
+  const sentinelRef = React.useRef<HTMLDivElement>(null);
+  const [stuck, setStuck] = React.useState(false);
+  React.useEffect(() => {
+    const el = sentinelRef.current;
+    if (!el || typeof IntersectionObserver === 'undefined') return;
+    const observer = new IntersectionObserver(([entry]) => setStuck(!entry.isIntersecting));
+    observer.observe(el);
+    return () => observer.disconnect();
+  }, []);
+
+  return (
+    <section aria-label={scene.title} className="flex flex-col gap-4">
+      <div ref={sentinelRef} aria-hidden className="-mb-4 h-px" />
+      <div
+        data-stuck={stuck}
+        className={cn(
+          'sticky top-0 z-30 -mx-3 flex flex-col gap-2 bg-card px-3 pb-3 pt-2 transition-shadow',
+          stuck && 'border-b border-border shadow-md',
+        )}
+      >
+        <div className="flex flex-wrap items-baseline gap-x-4 gap-y-1">
+          <h2 className="text-xl font-semibold text-foreground">{scene.title}</h2>
+          <p className="text-sm text-muted-foreground">
+            {scene.clips.length} clip{scene.clips.length === 1 ? '' : 's'} · {scene.shotCount} shot
+            {scene.shotCount === 1 ? '' : 's'} · about {formatClock(scene.totalSeconds)}
+          </p>
+          <span className="flex flex-wrap gap-1.5">
+            {statuses.map(([status, count]) => (
+              <Badge key={status} variant={BADGE_VARIANTS[status] ?? 'secondary'}>
+                {count} {status}
+              </Badge>
+            ))}
+          </span>
+        </div>
+        <Timeline scene={scene} />
+      </div>
+
+      <ol className="flex list-none flex-col gap-3 p-0">
+        {scene.clips.map((clip) => {
+          const first = shotNumber;
+          shotNumber += clip.shots.length;
+          return (
+            <ClipRow
+              key={clip.takeKey ?? clip.job.id}
+              clip={clip}
+              firstShotNumber={first}
+              onOpenJob={onOpenJob}
+              onChooseTake={onChooseTake}
+            />
+          );
+        })}
+      </ol>
+    </section>
+  );
+}
diff --git a/apps/fastvideo_studio/src/components/shell/Header.tsx b/apps/fastvideo_studio/src/components/shell/Header.tsx
index 2b1551cd4..443696e56 100644
--- a/apps/fastvideo_studio/src/components/shell/Header.tsx
+++ b/apps/fastvideo_studio/src/components/shell/Header.tsx
@@ -13,6 +13,7 @@ const TAB_TITLES: Record<string, string> = {
   '/distillation': 'Jobs',
   '/datasets': 'Datasets',
   '/gallery': 'Gallery',
+  '/scenes': 'Scenes',
   '/gpus': 'GPUs',
   '/settings': 'Settings',
 };
diff --git a/apps/fastvideo_studio/src/components/shell/PrimarySidebar.tsx b/apps/fastvideo_studio/src/components/shell/PrimarySidebar.tsx
index 2c277a82d..0bd73df41 100644
--- a/apps/fastvideo_studio/src/components/shell/PrimarySidebar.tsx
+++ b/apps/fastvideo_studio/src/components/shell/PrimarySidebar.tsx
@@ -150,6 +150,14 @@ export default function PrimarySidebar({
           >
             Datasets
           </Link>
+          <Link
+            href="/scenes"
+            aria-current={pathname === '/scenes' ? 'page' : undefined}
+            onClick={onMobileClose}
+            className={cn(TAB_BASE, pathname === '/scenes' && TAB_ACTIVE)}
+          >
+            Scenes
+          </Link>
           <Link
             href="/gallery"
             aria-current={pathname === '/gallery' ? 'page' : undefined}
diff --git a/apps/fastvideo_studio/src/components/ui/dialog.test.tsx b/apps/fastvideo_studio/src/components/ui/dialog.test.tsx
new file mode 100644
index 000000000..b66b038f4
--- /dev/null
+++ b/apps/fastvideo_studio/src/components/ui/dialog.test.tsx
@@ -0,0 +1,30 @@
+import { render, screen } from '@testing-library/react';
+import { describe, expect, it } from 'vitest';
+
+import { Dialog, DialogContent, DialogDescription, DialogTitle } from '@/components/ui/dialog';
+
+// The app header is `fixed` at z-[100] (components/shell/Header.tsx).
+const HEADER_Z = 100;
+
+function zOf(el: Element | null): number {
+  const m = el?.className.match(/\bz-\[(\d+)\]/);
+  return m ? Number(m[1]) : 0;
+}
+
+describe('Dialog stacking', () => {
+  it('renders its overlay and content above the fixed app header', () => {
+    render(
+      <Dialog open>
+        <DialogContent>
+          <DialogTitle>Tall dialog</DialogTitle>
+          <DialogDescription>Body</DialogDescription>
+        </DialogContent>
+      </Dialog>,
+    );
+    const content = screen.getByRole('dialog');
+    const overlay = document.querySelector('[data-state="open"].bg-black\\/70');
+
+    expect(zOf(content)).toBeGreaterThan(HEADER_Z);
+    expect(zOf(overlay)).toBeGreaterThan(HEADER_Z);
+  });
+});
diff --git a/apps/fastvideo_studio/src/components/ui/dialog.tsx b/apps/fastvideo_studio/src/components/ui/dialog.tsx
index 88eee674b..a92d00af4 100644
--- a/apps/fastvideo_studio/src/components/ui/dialog.tsx
+++ b/apps/fastvideo_studio/src/components/ui/dialog.tsx
@@ -11,6 +11,8 @@ const DialogTrigger = DialogPrimitive.Trigger;
 const DialogPortal = DialogPrimitive.Portal;
 const DialogClose = DialogPrimitive.Close;
 
+// Dialogs sit above the app header (z-[100], fixed). Below it, a tall dialog's
+// top edge slides under the header and, being fixed itself, can't be scrolled clear.
 const DialogOverlay = React.forwardRef<
   React.ElementRef<typeof DialogPrimitive.Overlay>,
   React.ComponentPropsWithoutRef<typeof DialogPrimitive.Overlay>
@@ -18,7 +20,7 @@ const DialogOverlay = React.forwardRef<
   <DialogPrimitive.Overlay
     ref={ref}
     className={cn(
-      'fixed inset-0 z-50 bg-black/70 backdrop-blur-sm data-[state=open]:animate-in data-[state=closed]:animate-out data-[state=closed]:fade-out-0 data-[state=open]:fade-in-0',
+      'fixed inset-0 z-[110] bg-black/70 backdrop-blur-sm data-[state=open]:animate-in data-[state=closed]:animate-out data-[state=closed]:fade-out-0 data-[state=open]:fade-in-0',
       className,
     )}
     {...props}
@@ -36,7 +38,7 @@ const DialogContent = React.forwardRef<
       ref={ref}
       data-modal=""
       className={cn(
-        'fixed left-1/2 top-1/2 z-50 grid w-full max-w-lg -translate-x-1/2 -translate-y-1/2 gap-4 rounded-2xl border border-border bg-card p-6 text-card-foreground shadow-lg duration-200 data-[state=open]:animate-in data-[state=closed]:animate-out data-[state=closed]:fade-out-0 data-[state=open]:fade-in-0',
+        'fixed left-1/2 top-1/2 z-[110] grid w-full max-w-lg -translate-x-1/2 -translate-y-1/2 gap-4 rounded-2xl border border-border bg-card p-6 text-card-foreground shadow-lg duration-200 data-[state=open]:animate-in data-[state=closed]:animate-out data-[state=closed]:fade-out-0 data-[state=open]:fade-in-0',
         className,
       )}
       {...props}
diff --git a/apps/fastvideo_studio/src/hooks/useUndoableState.test.ts b/apps/fastvideo_studio/src/hooks/useUndoableState.test.ts
new file mode 100644
index 000000000..f677b7616
--- /dev/null
+++ b/apps/fastvideo_studio/src/hooks/useUndoableState.test.ts
@@ -0,0 +1,140 @@
+import { act, renderHook } from '@testing-library/react';
+import { afterEach, beforeEach, describe, expect, it, vi } from 'vitest';
+
+import { COALESCE_MS, useUndoableState } from './useUndoableState';
+
+describe('useUndoableState', () => {
+  let now = 1_000_000;
+  beforeEach(() => {
+    now = 1_000_000;
+    vi.spyOn(Date, 'now').mockImplementation(() => now);
+  });
+  afterEach(() => vi.restoreAllMocks());
+
+  const tick = (ms: number) => {
+    now += ms;
+  };
+
+  it('makes each unkeyed change its own step', () => {
+    const { result } = renderHook(() => useUndoableState({ n: 0 }));
+    act(() => result.current.set({ n: 1 }));
+    act(() => result.current.set({ n: 2 }));
+    expect(result.current.state.n).toBe(2);
+
+    act(() => void result.current.undo());
+    expect(result.current.state.n).toBe(1);
+    act(() => void result.current.undo());
+    expect(result.current.state.n).toBe(0);
+    expect(result.current.canUndo).toBe(false);
+  });
+
+  it('merges same-key edits made close together into one step', () => {
+    const { result } = renderHook(() => useUndoableState('a'));
+    act(() => result.current.set('ab', 'text'));
+    tick(100);
+    act(() => result.current.set('abc', 'text'));
+    tick(100);
+    act(() => result.current.set('abcd', 'text'));
+
+    act(() => void result.current.undo());
+    expect(result.current.state).toBe('a');
+    expect(result.current.canUndo).toBe(false);
+  });
+
+  it('starts a new step after a pause, or when the key changes', () => {
+    const { result } = renderHook(() => useUndoableState('a'));
+    act(() => result.current.set('ab', 'text'));
+    tick(COALESCE_MS + 1);
+    act(() => result.current.set('abc', 'text'));
+    act(() => result.current.set('abcX', 'other'));
+
+    act(() => void result.current.undo());
+    expect(result.current.state).toBe('abc');
+    act(() => void result.current.undo());
+    expect(result.current.state).toBe('ab');
+    act(() => void result.current.undo());
+    expect(result.current.state).toBe('a');
+  });
+
+  it('never merges an unkeyed change into typing, or typing into it', () => {
+    const { result } = renderHook(() => useUndoableState('a'));
+    act(() => result.current.set('ab', 'text'));
+    act(() => result.current.set('AB'));
+    act(() => result.current.set('ABc', 'text'));
+
+    act(() => void result.current.undo());
+    expect(result.current.state).toBe('AB');
+    act(() => void result.current.undo());
+    expect(result.current.state).toBe('ab');
+  });
+
+  it('typing after an undo begins a fresh step instead of merging into the undone one', () => {
+    const { result } = renderHook(() => useUndoableState('a'));
+    act(() => result.current.set('ab', 'text'));
+    act(() => void result.current.undo());
+    act(() => result.current.set('ax', 'text'));
+    expect(result.current.canUndo).toBe(true);
+    act(() => void result.current.undo());
+    expect(result.current.state).toBe('a');
+  });
+
+  it('redoes, and a new change clears what could be redone', () => {
+    const { result } = renderHook(() => useUndoableState(0));
+    act(() => result.current.set(1));
+    act(() => result.current.set(2));
+    act(() => void result.current.undo());
+    act(() => void result.current.undo());
+    expect(result.current.canRedo).toBe(true);
+
+    act(() => void result.current.redo());
+    expect(result.current.state).toBe(1);
+    act(() => result.current.set(9));
+    expect(result.current.canRedo).toBe(false);
+    expect(result.current.redo()).toBe(false);
+  });
+
+  it('returns false when there is nothing to undo or redo', () => {
+    const { result } = renderHook(() => useUndoableState(0));
+    expect(result.current.undo()).toBe(false);
+    expect(result.current.redo()).toBe(false);
+  });
+
+  it('chains functional updates made in the same event', () => {
+    const { result } = renderHook(() => useUndoableState(0));
+    act(() => {
+      result.current.set((n) => n + 1);
+      result.current.set((n) => n + 1);
+    });
+    expect(result.current.state).toBe(2);
+    act(() => void result.current.undo());
+    expect(result.current.state).toBe(1);
+  });
+
+  it('ignores a set that produces the same value', () => {
+    const { result } = renderHook(() => useUndoableState('x'));
+    act(() => result.current.set('x'));
+    expect(result.current.canUndo).toBe(false);
+  });
+
+  it('reset replaces the value and drops all history', () => {
+    const { result } = renderHook(() => useUndoableState(0));
+    act(() => result.current.set(1));
+    act(() => void result.current.undo());
+    act(() => result.current.reset(7));
+    expect(result.current.state).toBe(7);
+    expect(result.current.canUndo).toBe(false);
+    expect(result.current.canRedo).toBe(false);
+  });
+
+  it('keeps at most 100 steps', () => {
+    const { result } = renderHook(() => useUndoableState(0));
+    for (let i = 1; i <= 130; i += 1) act(() => result.current.set(i));
+    let undone = 0;
+    while (result.current.canUndo) {
+      act(() => void result.current.undo());
+      undone += 1;
+    }
+    expect(undone).toBe(100);
+    expect(result.current.state).toBe(30);
+  });
+});
diff --git a/apps/fastvideo_studio/src/hooks/useUndoableState.ts b/apps/fastvideo_studio/src/hooks/useUndoableState.ts
new file mode 100644
index 000000000..3a7d974f4
--- /dev/null
+++ b/apps/fastvideo_studio/src/hooks/useUndoableState.ts
@@ -0,0 +1,86 @@
+'use client';
+
+import * as React from 'react';
+
+const MAX_HISTORY = 100;
+// Edits made under the same key, this close together, merge into one undo step.
+export const COALESCE_MS = 500;
+
+interface Snapshot<T> {
+  present: T;
+  past: T[];
+  future: T[];
+}
+
+/**
+ * Undo/redo over a whole state value. Every `set` is one undo step, except
+ * that consecutive sets with the same `key` inside COALESCE_MS merge -- pass a
+ * key for typing so a burst is one step, and none for discrete changes (moves,
+ * adds, selects) so each is its own. Values are treated as immutable, so
+ * snapshots share structure and are cheap to keep.
+ */
+export function useUndoableState<T>(initial: T) {
+  const [snapshot, setSnapshot] = React.useState<Snapshot<T>>({ present: initial, past: [], future: [] });
+  // Mirrors the snapshot synchronously so several sets in one event chain correctly.
+  const ref = React.useRef(snapshot);
+  const last = React.useRef<{ key: string | null; at: number }>({ key: null, at: 0 });
+
+  const apply = React.useCallback((next: Snapshot<T>) => {
+    ref.current = next;
+    setSnapshot(next);
+  }, []);
+
+  const set = React.useCallback(
+    (update: T | ((prev: T) => T), key: string | null = null) => {
+      const cur = ref.current;
+      const next = typeof update === 'function' ? (update as (prev: T) => T)(cur.present) : update;
+      if (Object.is(next, cur.present)) return;
+
+      const now = Date.now();
+      const merges = key !== null && last.current.key === key && now - last.current.at < COALESCE_MS;
+      last.current = { key, at: now };
+
+      if (merges) {
+        apply({ ...cur, present: next });
+      } else {
+        apply({ present: next, past: [...cur.past.slice(-(MAX_HISTORY - 1)), cur.present], future: [] });
+      }
+    },
+    [apply],
+  );
+
+  const undo = React.useCallback(() => {
+    const cur = ref.current;
+    if (cur.past.length === 0) return false;
+    last.current = { key: null, at: 0 };
+    apply({ present: cur.past[cur.past.length - 1], past: cur.past.slice(0, -1), future: [cur.present, ...cur.future] });
+    return true;
+  }, [apply]);
+
+  const redo = React.useCallback(() => {
+    const cur = ref.current;
+    if (cur.future.length === 0) return false;
+    last.current = { key: null, at: 0 };
+    apply({ present: cur.future[0], past: [...cur.past, cur.present], future: cur.future.slice(1) });
+    return true;
+  }, [apply]);
+
+  /** Replace the value and forget all history (e.g. when an editor reopens). */
+  const reset = React.useCallback(
+    (value: T) => {
+      last.current = { key: null, at: 0 };
+      apply({ present: value, past: [], future: [] });
+    },
+    [apply],
+  );
+
+  return {
+    state: snapshot.present,
+    set,
+    undo,
+    redo,
+    reset,
+    canUndo: snapshot.past.length > 0,
+    canRedo: snapshot.future.length > 0,
+  };
+}
diff --git a/apps/fastvideo_studio/src/hooks/useUndoableText.test.ts b/apps/fastvideo_studio/src/hooks/useUndoableText.test.ts
new file mode 100644
index 000000000..ff489f1ca
--- /dev/null
+++ b/apps/fastvideo_studio/src/hooks/useUndoableText.test.ts
@@ -0,0 +1,84 @@
+import { act, renderHook } from '@testing-library/react';
+import { afterEach, beforeEach, describe, expect, it, vi } from 'vitest';
+
+import { useUndoableText } from './useUndoableText';
+
+describe('useUndoableText', () => {
+  beforeEach(() => vi.useFakeTimers());
+  afterEach(() => vi.useRealTimers());
+
+  it('coalesces a typing burst into one undo step', () => {
+    const { result } = renderHook(() => useUndoableText('a'));
+    act(() => {
+      result.current.type('ab');
+      result.current.type('abc');
+    });
+    act(() => {
+      vi.advanceTimersByTime(600);
+    });
+    expect(result.current.canUndo).toBe(true);
+
+    let restored: string | null = null;
+    act(() => {
+      restored = result.current.undo();
+    });
+    expect(restored).toBe('a');
+    expect(result.current.text).toBe('a');
+    expect(result.current.canUndo).toBe(false);
+  });
+
+  it('gives a committed edit its own step and supports redo', () => {
+    const { result } = renderHook(() => useUndoableText('hi'));
+    act(() => result.current.commit('hi <pause>'));
+    expect(result.current.text).toBe('hi <pause>');
+
+    act(() => {
+      result.current.undo();
+    });
+    expect(result.current.text).toBe('hi');
+    expect(result.current.canRedo).toBe(true);
+
+    let redone: string | null = null;
+    act(() => {
+      redone = result.current.redo();
+    });
+    expect(redone).toBe('hi <pause>');
+    expect(result.current.text).toBe('hi <pause>');
+  });
+
+  it('flushes pending typing before a commit so undo steps stay separate', () => {
+    const { result } = renderHook(() => useUndoableText(''));
+    act(() => result.current.type('hello'));
+    act(() => result.current.commit('hello <pause>'));
+
+    act(() => {
+      result.current.undo();
+    });
+    expect(result.current.text).toBe('hello');
+    act(() => {
+      result.current.undo();
+    });
+    expect(result.current.text).toBe('');
+  });
+
+  it('returns null when there is nothing to undo or redo', () => {
+    const { result } = renderHook(() => useUndoableText('x'));
+    let undone: string | null = 'sentinel';
+    let redone: string | null = 'sentinel';
+    act(() => {
+      undone = result.current.undo();
+      redone = result.current.redo();
+    });
+    expect(undone).toBeNull();
+    expect(redone).toBeNull();
+  });
+
+  it('reset replaces the text and drops history', () => {
+    const { result } = renderHook(() => useUndoableText('a'));
+    act(() => result.current.commit('b'));
+    act(() => result.current.reset('fresh'));
+    expect(result.current.text).toBe('fresh');
+    expect(result.current.canUndo).toBe(false);
+    expect(result.current.canRedo).toBe(false);
+  });
+});
diff --git a/apps/fastvideo_studio/src/hooks/useUndoableText.ts b/apps/fastvideo_studio/src/hooks/useUndoableText.ts
new file mode 100644
index 000000000..4aa1e22ff
--- /dev/null
+++ b/apps/fastvideo_studio/src/hooks/useUndoableText.ts
@@ -0,0 +1,109 @@
+'use client';
+
+import * as React from 'react';
+
+const MAX_HISTORY = 100;
+// How long a pause in typing has to be before it becomes its own undo step.
+// Keeps undo from removing text one keystroke at a time while still giving
+// each tag/preset insertion (the actual ask) a guaranteed, immediate step.
+const TYPING_COALESCE_MS = 500;
+
+/**
+ * Text state with undo/redo. Typing is coalesced into bursts; discrete edits
+ * (`commit`, e.g. a tag insertion) always get their own step. `textRef`
+ * mirrors `text` synchronously because history bookkeeping needs the true
+ * current value the instant an edit happens, not whatever the last render
+ * closed over.
+ */
+export function useUndoableText(initial: string) {
+  const [text, setTextState] = React.useState(initial);
+  const textRef = React.useRef(initial);
+  const [past, setPast] = React.useState<string[]>([]);
+  const [future, setFuture] = React.useState<string[]>([]);
+  // Start of the in-progress typing burst that hasn't become a history step
+  // yet; flushed into `past` once typing pauses or a discrete edit happens.
+  const typingBaseRef = React.useRef<string | null>(null);
+  const typingTimerRef = React.useRef<ReturnType<typeof setTimeout> | null>(null);
+
+  const setText = React.useCallback((next: string) => {
+    textRef.current = next;
+    setTextState(next);
+  }, []);
+
+  const pushHistory = React.useCallback((previous: string) => {
+    setPast((p) => (p.length >= MAX_HISTORY ? [...p.slice(1), previous] : [...p, previous]));
+    setFuture([]);
+  }, []);
+
+  const clearTypingTimer = React.useCallback(() => {
+    if (typingTimerRef.current) {
+      clearTimeout(typingTimerRef.current);
+      typingTimerRef.current = null;
+    }
+  }, []);
+
+  const flushTypingCheckpoint = React.useCallback(() => {
+    clearTypingTimer();
+    if (typingBaseRef.current != null && typingBaseRef.current !== textRef.current) {
+      pushHistory(typingBaseRef.current);
+    }
+    typingBaseRef.current = null;
+  }, [clearTypingTimer, pushHistory]);
+
+  /** Free typing: coalesced into one undo step per burst. */
+  const type = React.useCallback(
+    (next: string) => {
+      if (typingBaseRef.current == null) typingBaseRef.current = textRef.current;
+      setText(next);
+      clearTypingTimer();
+      typingTimerRef.current = setTimeout(flushTypingCheckpoint, TYPING_COALESCE_MS);
+    },
+    [setText, clearTypingTimer, flushTypingCheckpoint],
+  );
+
+  /** Discrete edit (tag/preset insertion): always its own undo step. */
+  const commit = React.useCallback(
+    (next: string) => {
+      flushTypingCheckpoint();
+      pushHistory(textRef.current);
+      setText(next);
+    },
+    [flushTypingCheckpoint, pushHistory, setText],
+  );
+
+  /** Returns the restored text (so the caller can place the caret), or null if nothing to undo. */
+  function undo(): string | null {
+    flushTypingCheckpoint();
+    if (past.length === 0) return null;
+    const prev = past[past.length - 1];
+    setPast(past.slice(0, -1));
+    setFuture([...future, textRef.current]);
+    setText(prev);
+    return prev;
+  }
+
+  function redo(): string | null {
+    if (future.length === 0) return null;
+    const next = future[future.length - 1];
+    setFuture(future.slice(0, -1));
+    setPast([...past, textRef.current]);
+    setText(next);
+    return next;
+  }
+
+  /** Replace the text and drop all history (e.g. when an editor is reopened). */
+  const reset = React.useCallback(
+    (value: string) => {
+      setText(value);
+      setPast([]);
+      setFuture([]);
+      typingBaseRef.current = null;
+      clearTypingTimer();
+    },
+    [setText, clearTypingTimer],
+  );
+
+  React.useEffect(() => clearTypingTimer, [clearTypingTimer]);
+
+  return { text, type, commit, undo, redo, reset, canUndo: past.length > 0, canRedo: future.length > 0 };
+}
diff --git a/apps/fastvideo_studio/src/lib/h3DialogueEstimate.ts b/apps/fastvideo_studio/src/lib/h3DialogueEstimate.ts
new file mode 100644
index 000000000..ee84debb2
--- /dev/null
+++ b/apps/fastvideo_studio/src/lib/h3DialogueEstimate.ts
@@ -0,0 +1,171 @@
+/**
+ * Ballpark spoken-duration estimate for MiniMax-H3 dialogue text, plus the
+ * lexer/tree used to render a styled "performance preview" of the tags.
+ *
+ * Numbers here are heuristic (syllable counting + fixed per-tag seconds from
+ * `H3_TAGS[].seconds`), not measured against the model -- good enough to
+ * flag "this clip is clearly too long for one generation," calibrated
+ * against a real render via `calibrationFactor`.
+ */
+
+import { H3_TAGS, type H3Tag } from "@/lib/h3Tags";
+
+/** Bare tag name, e.g. "<long pause>" -> "long pause". */
+function bareName(bracketed: string): string {
+	return bracketed.replace(/[<>/]/g, "").trim();
+}
+
+const WRAP_TAG_NAMES = new Set(H3_TAGS.filter((t) => t.kind === "wrap").map((t) => bareName(t.open)));
+const TAG_SECONDS: Record<string, number> = Object.fromEntries(
+	H3_TAGS.map((t) => [bareName(t.open), t.seconds ?? 0]),
+);
+/** Bare tag name -> its registry entry, so callers (the preview renderer) can look up category/description. */
+export const TAG_BY_NAME: Record<string, H3Tag> = Object.fromEntries(H3_TAGS.map((t) => [bareName(t.open), t]));
+
+/* -------------------------------------------------------------------- */
+/*  Lexer / tree — mirrors the tag registry so preview & estimate always  */
+/*  agree with what the console/editor can actually insert.               */
+/* -------------------------------------------------------------------- */
+
+export type LexToken =
+	| { t: "text"; v: string }
+	| { t: "point"; name: string }
+	| { t: "open"; name: string }
+	| { t: "close"; name: string };
+
+const TAG_RE = /<(\/?)([a-zA-Z ]+?)>/g;
+
+export function lexDialogue(text: string): LexToken[] {
+	const tokens: LexToken[] = [];
+	let last = 0;
+	let m: RegExpExecArray | null;
+	TAG_RE.lastIndex = 0;
+	while ((m = TAG_RE.exec(text))) {
+		if (m.index > last) tokens.push({ t: "text", v: text.slice(last, m.index) });
+		const closing = m[1] === "/";
+		const name = m[2].trim();
+		if (WRAP_TAG_NAMES.has(name)) tokens.push({ t: closing ? "close" : "open", name });
+		else if (!closing) tokens.push({ t: "point", name }); // stray/unknown closers are ignored
+		last = TAG_RE.lastIndex;
+	}
+	if (last < text.length) tokens.push({ t: "text", v: text.slice(last) });
+	return tokens;
+}
+
+export interface DialogueTextNode {
+	t: "text";
+	v: string;
+}
+export interface DialoguePointNode {
+	t: "point";
+	name: string;
+}
+export interface DialogueSpanNode {
+	t: "span";
+	name: string;
+	children: DialogueNode[];
+}
+export type DialogueNode = DialogueTextNode | DialoguePointNode | DialogueSpanNode;
+
+export function dialogueTree(text: string): DialogueNode[] {
+	const root: { children: DialogueNode[] } = { children: [] };
+	const stack: { children: DialogueNode[] }[] = [root];
+	for (const tk of lexDialogue(text)) {
+		const top = stack[stack.length - 1];
+		if (tk.t === "text") top.children.push({ t: "text", v: tk.v });
+		else if (tk.t === "point") top.children.push({ t: "point", name: tk.name });
+		else if (tk.t === "open") {
+			const node: DialogueSpanNode = { t: "span", name: tk.name, children: [] };
+			top.children.push(node);
+			stack.push(node);
+		} else if (tk.t === "close" && stack.length > 1) stack.pop();
+	}
+	return root.children;
+}
+
+/* -------------------------------------------------------------------- */
+/*  Duration estimate                                                    */
+/* -------------------------------------------------------------------- */
+
+export const SPEECH_RATES = { Measured: 3.2, Natural: 4.0, Quick: 4.8 } as const; // syllables/sec
+export type SpeechRate = keyof typeof SPEECH_RATES;
+
+export function syllables(word: string): number {
+	let w = word.toLowerCase().replace(/[^a-z]/g, "");
+	if (!w) return 0;
+	if (w.length <= 3) return 1;
+	w = w.replace(/(?:[^laeiouy]es|ed|[^laeiouy]e)$/, "");
+	w = w.replace(/^y/, "");
+	const m = w.match(/[aeiouy]{1,2}/g);
+	return m ? m.length : 1;
+}
+
+export interface DurationEstimate {
+	total: number;
+	speech: number;
+	punct: number;
+	tags: number;
+	syl: number;
+	words: number;
+	low: number;
+	high: number;
+}
+
+export function estimateDuration(text: string, rate: number, calib: number): DurationEstimate {
+	const spoken = text.replace(/<\/?[a-zA-Z ]+?>/g, " ");
+	const words = spoken.split(/\s+/).filter(Boolean);
+	const syl = words.reduce((n, w) => n + syllables(w), 0);
+	const speech = syl / rate;
+
+	const sentence = (spoken.match(/[.!?]/g) || []).length * 0.35;
+	const commas = (spoken.match(/[,;:]/g) || []).length * 0.18;
+	const ellipsis = (spoken.match(/\.\.\.|…/g) || []).length * 0.5;
+	const dashes = (spoken.match(/—/g) || []).length * 0.2;
+	const punct = sentence + commas + ellipsis + dashes;
+
+	let tags = 0;
+	const re = /<(\/?)([a-zA-Z ]+?)>/g;
+	let m: RegExpExecArray | null;
+	while ((m = re.exec(text))) {
+		if (m[1] === "/") continue;
+		tags += TAG_SECONDS[m[2].trim()] ?? 0;
+	}
+
+	const total = (speech + punct + tags) * calib;
+	return {
+		total,
+		speech: speech * calib,
+		punct: punct * calib,
+		tags: tags * calib,
+		syl,
+		words: words.length,
+		low: total * 0.86,
+		high: total * 1.16,
+	};
+}
+
+/** Find the best sentence/pause boundary at which to split text that runs past `ceiling` seconds. */
+export function suggestSplit(text: string, rate: number, calib: number, ceiling: number): number {
+	const re = /(<\/?[a-zA-Z ]+?>)|([.!?]+)|([,;:]+)|(\s+)|([^\s<.!?,;:]+)/g;
+	let cum = 0;
+	let boundary = 0;
+	let m: RegExpExecArray | null;
+	const budget = ceiling - 1.5;
+	while ((m = re.exec(text))) {
+		if (m[1]) {
+			const closing = m[1].startsWith("</");
+			const name = bareName(m[1]);
+			if (!closing) cum += (TAG_SECONDS[name] ?? 0) * calib;
+			if (name === "pause" || name === "long pause") boundary = re.lastIndex;
+		} else if (m[2]) {
+			cum += 0.35 * calib;
+			boundary = re.lastIndex;
+		} else if (m[3]) {
+			cum += 0.18 * calib;
+		} else if (m[5]) {
+			cum += (syllables(m[5]) / rate) * calib;
+		}
+		if (cum > budget && boundary > 0) return boundary;
+	}
+	return boundary > 0 ? boundary : -1;
+}
diff --git a/apps/fastvideo_studio/src/lib/h3PromptFields.test.ts b/apps/fastvideo_studio/src/lib/h3PromptFields.test.ts
new file mode 100644
index 000000000..6a4577ea0
--- /dev/null
+++ b/apps/fastvideo_studio/src/lib/h3PromptFields.test.ts
@@ -0,0 +1,87 @@
+import { describe, expect, it } from "vitest";
+import { readField, shotTarget, writeField, writeShotTarget } from "@/lib/h3PromptFields";
+
+const BASE = [
+	"For the target video, <Picture 1> is fully referenced.",
+	"",
+	"integrated_multimodal_description: [Shot 1] A close-up. He (S1) says: <d>[English] Hi.</d> <cutoff>",
+	"",
+	"overall_soundscape: Room tone.",
+	"",
+	"non_diegetic_music: N/A",
+].join("\n");
+
+const SECTIONS = [
+	"subject_definitions:",
+	"<Subject 1> is the courier.",
+	"",
+	"retention_analysis:",
+	"<Subject 1> (appears in [Shot 1]): fully_preserved - x.",
+	"",
+	"detailed_description:",
+	"[Shot 1] Wide shot. A rooftop.",
+	"[Shot 2] At 00:02.000, the shot cuts to a close-up shot. A face.",
+	"",
+	"overall_soundscape:",
+	"Rain.",
+].join("\n");
+
+describe("readField / writeField", () => {
+	it("reads a field whose text follows the heading on the same line", () => {
+		expect(readField(BASE, "integrated_multimodal_description")).toBe(
+			"[Shot 1] A close-up. He (S1) says: <d>[English] Hi.</d> <cutoff>",
+		);
+	});
+
+	it("reads a field whose text starts on the line after the heading", () => {
+		expect(readField(SECTIONS, "detailed_description")).toBe(
+			"[Shot 1] Wide shot. A rooftop.\n[Shot 2] At 00:02.000, the shot cuts to a close-up shot. A face.",
+		);
+		expect(readField(SECTIONS, "subject_definitions")).toBe("<Subject 1> is the courier.");
+	});
+
+	it("returns null for a field that isn't there", () => {
+		expect(readField(BASE, "detailed_description")).toBeNull();
+	});
+
+	it("rewrites only that field, leaving the rest byte-identical (inline style)", () => {
+		const out = writeField(BASE, "integrated_multimodal_description", "[Shot 1] New.");
+		expect(out).toBe(BASE.replace("[Shot 1] A close-up. He (S1) says: <d>[English] Hi.</d> <cutoff>", "[Shot 1] New."));
+	});
+
+	it("rewrites only that field (own-line style)", () => {
+		const out = writeField(SECTIONS, "detailed_description", "[Shot 1] Only.");
+		expect(out).toContain("detailed_description:\n[Shot 1] Only.\n\noverall_soundscape:\nRain.");
+		expect(out.startsWith("subject_definitions:\n<Subject 1> is the courier.")).toBe(true);
+	});
+
+	it("writing a field's own text back is a no-op", () => {
+		for (const [prompt, field] of [
+			[BASE, "integrated_multimodal_description"],
+			[SECTIONS, "detailed_description"],
+			[SECTIONS, "retention_analysis"],
+		] as const) {
+			expect(writeField(prompt, field, readField(prompt, field) as string)).toBe(prompt);
+		}
+	});
+
+	it("ignores a heading that isn't at the start of a line", () => {
+		expect(readField("see detailed_description: not a header", "detailed_description")).toBeNull();
+	});
+});
+
+describe("shotTarget", () => {
+	it("prefers integrated_multimodal_description", () => {
+		expect(shotTarget(BASE).field).toBe("integrated_multimodal_description");
+	});
+
+	it("falls back to detailed_description", () => {
+		expect(shotTarget(SECTIONS).field).toBe("detailed_description");
+	});
+
+	it("treats a plain prompt as the shot text itself", () => {
+		const t = shotTarget("A toy car drives into a plush dog.");
+		expect(t).toEqual({ field: null, body: "A toy car drives into a plush dog." });
+		expect(writeShotTarget("A toy car drives into a plush dog.", t, "[Shot 1] X.")).toBe("[Shot 1] X.");
+	});
+});
diff --git a/apps/fastvideo_studio/src/lib/h3PromptFields.ts b/apps/fastvideo_studio/src/lib/h3PromptFields.ts
new file mode 100644
index 000000000..d8330d525
--- /dev/null
+++ b/apps/fastvideo_studio/src/lib/h3PromptFields.ts
@@ -0,0 +1,77 @@
+/**
+ * Locate and rewrite a single named field inside a raw MiniMax-H3 prompt,
+ * leaving every other character of the prompt untouched. Handles both header
+ * styles seen in real prompts: the heading on its own line (six-section
+ * reference format) and the heading followed by its text on the same line
+ * (`integrated_multimodal_description: [Shot 1] ...`, the base format).
+ */
+
+const FIELD_NAMES = [
+	"subject_definitions",
+	"summary",
+	"retention_analysis",
+	"detailed_description",
+	"integrated_multimodal_description",
+	"overall_soundscape",
+	"non_diegetic_music",
+] as const;
+
+export type H3PromptFieldName = (typeof FIELD_NAMES)[number];
+
+const HEADER_RE = new RegExp(`^(${FIELD_NAMES.join("|")}):[ \\t]*`, "gm");
+
+interface Span {
+	start: number;
+	end: number;
+}
+
+function locate(prompt: string, field: H3PromptFieldName): Span | null {
+	const headers = [...prompt.matchAll(HEADER_RE)];
+	const i = headers.findIndex((h) => h[1] === field);
+	if (i < 0) return null;
+
+	const header = headers[i];
+	let start = (header.index ?? 0) + header[0].length;
+	// Own-line style: nothing follows the colon, so the body starts on the next line.
+	const newline = prompt.slice(start).match(/^\r?\n/);
+	if (newline) start += newline[0].length;
+
+	const nextHeader = i + 1 < headers.length ? (headers[i + 1].index ?? prompt.length) : prompt.length;
+	const end = start + Math.max(0, prompt.slice(start, nextHeader).trimEnd().length);
+	return { start, end: Math.max(start, end) };
+}
+
+/** The field's text exactly as written, or null if the prompt has no such field. */
+export function readField(prompt: string, field: H3PromptFieldName): string | null {
+	const span = locate(prompt, field);
+	return span ? prompt.slice(span.start, span.end) : null;
+}
+
+/** Replace only the field's text. Returns the prompt unchanged if the field isn't there. */
+export function writeField(prompt: string, field: H3PromptFieldName, body: string): string {
+	const span = locate(prompt, field);
+	return span ? prompt.slice(0, span.start) + body + prompt.slice(span.end) : prompt;
+}
+
+export interface ShotTarget {
+	/** the field the [Shot N] blocks live in, or null when the whole prompt is the shot text */
+	field: "integrated_multimodal_description" | "detailed_description" | null;
+	body: string;
+}
+
+/**
+ * Where a prompt's shots live: the base format's integrated_multimodal_description,
+ * else the reference format's detailed_description, else the prompt as a whole.
+ */
+export function shotTarget(prompt: string): ShotTarget {
+	for (const field of ["integrated_multimodal_description", "detailed_description"] as const) {
+		const body = readField(prompt, field);
+		if (body !== null) return { field, body };
+	}
+	return { field: null, body: prompt };
+}
+
+/** Write shot text back to wherever shotTarget found it. */
+export function writeShotTarget(prompt: string, target: ShotTarget, body: string): string {
+	return target.field ? writeField(prompt, target.field, body) : body;
+}
diff --git a/apps/fastvideo_studio/src/lib/h3ScriptImport.test.ts b/apps/fastvideo_studio/src/lib/h3ScriptImport.test.ts
new file mode 100644
index 000000000..de7050590
--- /dev/null
+++ b/apps/fastvideo_studio/src/lib/h3ScriptImport.test.ts
@@ -0,0 +1,238 @@
+import { describe, expect, it } from "vitest";
+import {
+	buildClips,
+	convertDirections,
+	parseScript,
+	scriptSpeakers,
+	type ImportSpeaker,
+} from "@/lib/h3ScriptImport";
+import { detailedDescriptionFromShots, shotLines, shotProse } from "@/lib/h3Shots";
+
+// Original test material -- not from any real screenplay.
+const SCRIPT = [
+	"INT. HARBOR OFFICE - DAY",
+	"",
+	"CAPTAIN: Sit down. (EXHALES) We need to talk about the ledger.",
+	"",
+	"CLERK: Yes, sir.",
+	"",
+	"CAPTAIN: Three crates are missing, and nobody on this dock saw a thing.",
+	"Not one of you. That does not happen by accident.",
+	"",
+	"CLERK: I counted twice.",
+	"",
+	"CAPTAIN: (WHISPERING) Then count a third time.",
+	"",
+	"CLERK: (TAPPING TABLE) Right away.",
+].join("\n");
+
+const SPEAKERS: Record<string, ImportSpeaker> = {
+	CAPTAIN: { subject: "<Subject 1>", tag: "S1" },
+	CLERK: { subject: "<Subject 2>", tag: "S2" },
+};
+
+describe("parseScript", () => {
+	it("reads NAME: line turns and ignores headings between turns", () => {
+		const turns = parseScript(SCRIPT);
+		expect(turns.map((t) => t.speaker)).toEqual(["CAPTAIN", "CLERK", "CAPTAIN", "CLERK", "CAPTAIN", "CLERK"]);
+		expect(turns.some((t) => t.text.includes("HARBOR"))).toBe(false);
+	});
+
+	it("joins continuation lines directly under a turn", () => {
+		const turns = parseScript(SCRIPT);
+		expect(turns[2].text).toBe(
+			"Three crates are missing, and nobody on this dock saw a thing. Not one of you. That does not happen by accident.",
+		);
+	});
+
+	it("tolerates markdown bold speaker names", () => {
+		expect(parseScript("**MARK:** hello there")).toEqual([{ speaker: "MARK", text: "hello there" }]);
+	});
+
+	it("lists speakers in order of first appearance", () => {
+		expect(scriptSpeakers(parseScript(SCRIPT))).toEqual(["CAPTAIN", "CLERK"]);
+	});
+
+	it("returns nothing for text with no speaker turns", () => {
+		expect(parseScript("just some prose without turns")).toEqual([]);
+	});
+});
+
+describe("convertDirections", () => {
+	it("turns known parentheticals into speech tags in place", () => {
+		expect(convertDirections("Sit down. (EXHALES) We talk.").text).toBe("Sit down. <exhale> We talk.");
+		expect(convertDirections("Well (BEAT) fine.").text).toBe("Well <pause> fine.");
+	});
+
+	it("wraps the line for a whispering direction", () => {
+		expect(convertDirections("(WHISPERING) Keep it down.").text).toBe("<whisper>Keep it down.</whisper>");
+	});
+
+	it("keeps unmapped directions out of the dialogue but reports them", () => {
+		const r = convertDirections("(TAPPING TABLE) Right away.");
+		expect(r.text).toBe("Right away.");
+		expect(r.directions).toEqual(["tapping table"]);
+	});
+});
+
+describe("buildClips", () => {
+	it("opens the first clip with an establishing shot", () => {
+		const [clip] = buildClips(parseScript(SCRIPT), { speakers: SPEAKERS, targetSeconds: 60 });
+		expect(clip.shots[0].type).toBe("establishing");
+		expect(shotLines(clip.shots[0])).toHaveLength(0);
+		expect(clip.shots[0].subjectIds).toEqual(["<Subject 1>", "<Subject 2>"]);
+	});
+
+	it("folds a short reaction into the shot it reacts to", () => {
+		const [clip] = buildClips(parseScript(SCRIPT), { speakers: SPEAKERS, targetSeconds: 60 });
+		const first = clip.shots[1];
+		expect(shotLines(first).map((l) => [l.speaker, l.text])).toEqual([
+			["S1", "Sit down. <exhale> We need to talk about the ledger."],
+			["S2", "Yes, sir."],
+		]);
+		expect(first.type).toBe("medium");
+		expect(first.subjectIds).toEqual(["<Subject 1>", "<Subject 2>"]);
+	});
+
+	it("uses a close-up when only one person speaks in a shot", () => {
+		const [clip] = buildClips([{ speaker: "CAPTAIN", text: "Nobody leaves until the count is done." }], {
+			speakers: SPEAKERS,
+			targetSeconds: 60,
+		});
+		const shot = clip.shots[1];
+		expect(shot.type).toBe("close-up");
+		expect(shot.subjectIds).toEqual(["<Subject 1>"]);
+		expect(shotProse(shot)).toBe("Close on <Subject 1>, speaking.");
+	});
+
+	it("records unmapped stage directions on the shot they belong to", () => {
+		const [clip] = buildClips(parseScript(SCRIPT), { speakers: SPEAKERS, targetSeconds: 60 });
+		const last = clip.shots[clip.shots.length - 1];
+		expect(shotLines(last).map((l) => l.text)).toEqual(["<whisper>Then count a third time.</whisper>", "Right away."]);
+		expect(shotProse(last)).toContain("Stage direction: tapping table.");
+	});
+
+	it("splits a long scene into several clips, each with dialogue", () => {
+		const clips = buildClips(parseScript(SCRIPT), { speakers: SPEAKERS, targetSeconds: 5 });
+		expect(clips.length).toBeGreaterThan(1);
+		for (const c of clips) expect(c.shots.some((s) => shotLines(s).length > 0)).toBe(true);
+		expect(clips[1].shots[0].type).not.toBe("establishing");
+	});
+
+	it("keeps every spoken word across clips", () => {
+		const clips = buildClips(parseScript(SCRIPT), { speakers: SPEAKERS, targetSeconds: 5 });
+		const all = clips.flatMap((c) => c.shots.flatMap((s) => shotLines(s).map((l) => l.text))).join(" ");
+		for (const word of ["ledger", "crates", "accident", "counted", "third", "Right away"]) {
+			expect(all).toContain(word);
+		}
+	});
+
+	it("serializes to valid [Shot N] text with speaker-tagged dialogue", () => {
+		const [clip] = buildClips(parseScript(SCRIPT), { speakers: SPEAKERS, targetSeconds: 60 });
+		const text = detailedDescriptionFromShots(clip.shots);
+		expect(text).toContain("[Shot 1] Establishing shot.");
+		expect(text).toContain("(S1) <d>[English] Sit down. <exhale> We need to talk about the ledger.</d>");
+		expect(text).toContain("(S2) <d>[English] Yes, sir.</d>");
+	});
+
+	describe("speaker tags per clip", () => {
+		it("numbers speakers by who speaks first in each clip, not across the whole script", () => {
+			const turns = [
+				{ speaker: "CAPTAIN", text: "Every one of these lines is long enough to stand as its own turn today." },
+				{ speaker: "CLERK", text: "And every reply here is long enough to stand as its own turn as well." },
+				{ speaker: "CLERK", text: "The clerk speaks again first when the next clip begins, with a full sentence." },
+				{ speaker: "CAPTAIN", text: "And then the captain answers the clerk with another complete sentence here." },
+			];
+			const clips = buildClips(turns, { speakers: SPEAKERS, targetSeconds: 8 });
+			expect(clips.length).toBeGreaterThan(1);
+			for (const c of clips) {
+				const first = c.shots.flatMap((s) => shotLines(s))[0];
+				expect(first.speaker).toBe("S1");
+			}
+			// a clip that opens on the clerk maps the clerk to S1 and the captain to S2
+			const clerkFirst = clips.find((c) => shotLines(c.shots.find((s) => shotLines(s).length)!)[0].text.startsWith("The clerk"));
+			const speakers = clerkFirst!.shots.flatMap((s) => shotLines(s).map((l) => l.speaker));
+			expect(new Set(speakers).size).toBeLessThanOrEqual(2);
+		});
+
+		it("leaves custom speaker tags alone", () => {
+			const clips = buildClips(
+				[{ speaker: "CLERK", text: "A line for the clerk that is long enough to stand alone here." }],
+				{ speakers: { CLERK: { subject: "<Subject 2>", tag: "CLERK" }, CAPTAIN: SPEAKERS.CAPTAIN }, targetSeconds: 11 },
+			);
+			expect(shotLines(clips[0].shots[1])[0].speaker).toBe("CLERK");
+		});
+	});
+
+	describe("clip boundaries", () => {
+		const longSpeech = Array.from({ length: 12 }, (_, i) => `This is sentence number ${i + 1} of a long speech.`).join(" ");
+
+		it("keeps a long turn as one continuous shot when a clip can hold it", () => {
+			const clips = buildClips([{ speaker: "CAPTAIN", text: longSpeech }], { speakers: SPEAKERS, targetSeconds: 60 });
+			expect(clips).toHaveLength(1);
+			const dialogueShots = clips[0].shots.filter((s) => shotLines(s).length > 0);
+			expect(dialogueShots).toHaveLength(1);
+			expect(shotLines(dialogueShots[0])[0].text).toBe(longSpeech);
+		});
+
+		it("only splits inside a shot when the turn is longer than a whole clip, and marks the follow-on clips", () => {
+			const clips = buildClips([{ speaker: "CAPTAIN", text: longSpeech }], { speakers: SPEAKERS, targetSeconds: 11 });
+			expect(clips.length).toBeGreaterThan(1);
+			expect(clips[0].continuation).toBe(false);
+			for (const c of clips.slice(1)) {
+				expect(c.continuation).toBe(true);
+				expect(shotProse(c.shots[0])).toContain("Continuing the same shot");
+				expect(c.shots[0].type).toBe("close-up");
+			}
+			const spoken = clips.flatMap((c) => c.shots.flatMap((s) => shotLines(s).map((l) => l.text))).join(" ");
+			expect(spoken).toBe(longSpeech);
+		});
+
+		it("marks speech that runs past the end of a clip with <cutoff>, and only there", () => {
+			const clips = buildClips([{ speaker: "CAPTAIN", text: longSpeech }], { speakers: SPEAKERS, targetSeconds: 11 });
+			const endings = clips.map((c) => detailedDescriptionFromShots(c.shots));
+			for (const text of endings.slice(0, -1)) expect(text).toMatch(/<\/d> <cutoff>$/);
+			expect(endings[endings.length - 1]).not.toContain("<cutoff>");
+			// the marker is outside the spoken text, so it never leaks into the dialogue itself
+			const spoken = clips.flatMap((c) => c.shots.flatMap((s) => shotLines(s).map((l) => l.text))).join(" ");
+			expect(spoken).not.toContain("cutoff");
+		});
+
+		it("breaks between separate turns at a cut, not as a continuation", () => {
+			const turns = Array.from({ length: 6 }, () => ({
+				speaker: "CAPTAIN",
+				text: "Every one of these turns is a complete thought that takes a while to say aloud.",
+			}));
+			const clips = buildClips(turns, { speakers: SPEAKERS, targetSeconds: 11 });
+			expect(clips.length).toBeGreaterThan(1);
+			expect(clips.every((c) => !c.continuation)).toBe(true);
+		});
+
+		it("re-wraps a whispered line on every piece so no tag is left unmatched", () => {
+			const clips = buildClips([{ speaker: "CAPTAIN", text: `(WHISPERING) ${longSpeech}` }], {
+				speakers: SPEAKERS,
+				targetSeconds: 11,
+			});
+			const texts = clips.flatMap((c) => c.shots.flatMap((s) => shotLines(s).map((l) => l.text)));
+			expect(texts.length).toBeGreaterThan(1);
+			for (const t of texts) {
+				expect(t.startsWith("<whisper>")).toBe(true);
+				expect(t.endsWith("</whisper>")).toBe(true);
+			}
+		});
+
+		it("puts a trailing reaction with the last piece of a split speech", () => {
+			const clips = buildClips(
+				[
+					{ speaker: "CAPTAIN", text: longSpeech },
+					{ speaker: "CLERK", text: "Understood." },
+				],
+				{ speakers: SPEAKERS, targetSeconds: 11 },
+			);
+			const lastShot = clips[clips.length - 1].shots.at(-1)!;
+			expect(shotLines(lastShot).map((l) => l.speaker)).toEqual(["S1", "S2"]);
+			const earlier = clips.slice(0, -1).flatMap((c) => c.shots.flatMap((s) => shotLines(s).map((l) => l.speaker)));
+			expect(earlier.every((sp) => sp === "S1")).toBe(true);
+		});
+	});
+});
diff --git a/apps/fastvideo_studio/src/lib/h3ScriptImport.ts b/apps/fastvideo_studio/src/lib/h3ScriptImport.ts
new file mode 100644
index 000000000..823e993d8
--- /dev/null
+++ b/apps/fastvideo_studio/src/lib/h3ScriptImport.ts
@@ -0,0 +1,332 @@
+/**
+ * Turns a plain dialogue script ("NAME: line" turns) into clip-sized shot
+ * lists for the shot editor. Mechanical work only -- speakers, dialogue,
+ * stage-direction-to-tag mapping, grouping turns into shots, and splitting
+ * the whole scene into clips that fit one generation. Camera choices are
+ * defaults for the user to refine; a script says nothing about them.
+ */
+
+import { SPEECH_RATES, estimateDuration } from "@/lib/h3DialogueEstimate";
+import { estimateSpokenSeconds, makeShot, newLine, newTextPart, shotLines, type Shot } from "@/lib/h3Shots";
+import { H3_TAGS } from "@/lib/h3Tags";
+
+export interface ScriptTurn {
+	speaker: string;
+	text: string;
+}
+
+// "MARK: ..." -- 1-3 capitalized words, so prose like "note: ..." or scene
+// headings without a colon are never mistaken for a speaker.
+const SPEAKER_LINE_RE = /^([A-Z][A-Za-z0-9.'_-]*(?: [A-Z][A-Za-z0-9.'_-]*){0,2}):\s*(.*)$/;
+
+function stripMarkdown(line: string): string {
+	return line
+		.replace(/^\s*(?:[>*-]\s+)+/, "")
+		.replace(/\*\*|__/g, "")
+		.trim();
+}
+
+/**
+ * A speaker line starts a turn; non-blank lines directly after it continue
+ * that turn. A blank line ends the turn, so scene headings and action lines
+ * between turns are ignored rather than glued onto someone's dialogue.
+ */
+export function parseScript(script: string): ScriptTurn[] {
+	const turns: ScriptTurn[] = [];
+	let continuing = false;
+	for (const raw of script.split(/\r?\n/)) {
+		const line = stripMarkdown(raw);
+		if (!line) {
+			continuing = false;
+			continue;
+		}
+		const m = line.match(SPEAKER_LINE_RE);
+		if (m) {
+			turns.push({ speaker: m[1], text: m[2].trim() });
+			continuing = true;
+		} else if (continuing && turns.length > 0) {
+			const last = turns[turns.length - 1];
+			last.text = `${last.text} ${line}`.trim();
+		}
+	}
+	return turns.filter((t) => t.text);
+}
+
+/** Speaker names in order of first appearance. */
+export function scriptSpeakers(turns: ScriptTurn[]): string[] {
+	const seen: string[] = [];
+	for (const t of turns) if (!seen.includes(t.speaker)) seen.push(t.speaker);
+	return seen;
+}
+
+/* ---------------------------------------------------------------- */
+/*  Stage directions -> speech tags                                   */
+/* ---------------------------------------------------------------- */
+
+const TAG_BY_ID = Object.fromEntries(H3_TAGS.map((t) => [t.id, t]));
+
+// Order matters: the more specific phrase must be tried before its substring.
+const DIRECTION_TAGS: [RegExp, string][] = [
+	[/\blong pause\b/, "long-pause"],
+	[/\bcatch(?:es)? (?:his |her |their )?breath\b/, "catches-breath"],
+	[/\bexhal/, "exhale"],
+	[/\binhal/, "inhale"],
+	[/\bsigh/, "sighs"],
+	[/\blaugh/, "laughs"],
+	[/\bchuckl/, "chuckle"],
+	[/\bgasp/, "gasp"],
+	[/\bcough/, "coughs"],
+	[/\bclears? (?:his |her |their )?throat\b/, "clears-throat"],
+	[/\bsniff/, "sniff"],
+	[/\b(?:pause|beat)\b/, "pause"],
+	[/\bstutter|\bstammer/, "stutter"],
+	[/\bbreath/, "breath"],
+];
+
+export interface ConvertedLine {
+	text: string;
+	/** parentheticals with no matching tag, kept as stage directions for the shot description */
+	directions: string[];
+}
+
+/** Rewrites "(EXHALES)"-style parentheticals as speech tags; "(WHISPERING)" wraps the whole line. */
+export function convertDirections(text: string): ConvertedLine {
+	const directions: string[] = [];
+	let wrapWhisper = false;
+
+	let out = text.replace(/\(([^)]*)\)/g, (_m, inner: string) => {
+		const key = inner.trim().toLowerCase();
+		if (/\bwhisper/.test(key)) {
+			wrapWhisper = true;
+			return " ";
+		}
+		for (const [re, id] of DIRECTION_TAGS) {
+			const tag = TAG_BY_ID[id];
+			if (tag && tag.kind === "point" && re.test(key)) return ` ${tag.open} `;
+		}
+		if (key) directions.push(key);
+		return " ";
+	});
+
+	out = out
+		.replace(/\s+/g, " ")
+		.replace(/\s+([.,!?…])/g, "$1")
+		.trim();
+
+	const whisper = TAG_BY_ID.whisper;
+	if (wrapWhisper && whisper?.close && out) out = `${whisper.open}${out}${whisper.close}`;
+	return { text: out, directions };
+}
+
+/* ---------------------------------------------------------------- */
+/*  Shots and clips                                                   */
+/* ---------------------------------------------------------------- */
+
+export interface ImportSpeaker {
+	/** <Subject N> label this speaker is, or null when they aren't a tracked subject */
+	subject: string | null;
+	/** speaker tag written as "(S1)" */
+	tag: string;
+}
+
+export interface ImportOptions {
+	speakers: Record<string, ImportSpeaker>;
+	/** rough length one clip should fit, in seconds */
+	targetSeconds: number;
+}
+
+export interface ImportClip {
+	shots: Shot[];
+	/**
+	 * True when this clip picks up in the middle of the previous clip's last
+	 * shot (a speech too long for one generation). Its first shot continues
+	 * that one instead of cutting to something new, so it should be generated
+	 * from the previous clip's last frame. False means the boundary is a cut.
+	 */
+	continuation: boolean;
+}
+
+const ESTABLISHING_SECONDS = 1.5;
+const REACTION_MAX_WORDS = 4;
+// Space left in a clip is only worth starting a forced mid-shot split in if
+// it can hold at least this much of the speech.
+const MIN_PART_SECONDS = 3;
+
+function spokenSeconds(text: string): number {
+	return estimateDuration(text, SPEECH_RATES.Natural, 1).total;
+}
+
+function wordCount(text: string): number {
+	return (text.replace(/<[^>]*>/g, " ").match(/[\p{L}\p{N}'’-]+/gu) ?? []).length;
+}
+
+/**
+ * Split one over-long line at sentence boundaries into parts that fit the
+ * given budgets (first part, then every later part). An outer wrapping tag
+ * like <whisper>...</whisper> is re-applied to each part so none is left
+ * with an unmatched tag.
+ */
+function splitToBudgets(text: string, firstBudget: number, laterBudget: number): string[] {
+	const wrap = text.match(/^<([a-z ]+)>([\s\S]*)<\/\1>$/);
+	const inner = wrap ? wrap[2] : text;
+	const sentences = inner.match(/[^.!?…]+(?:[.!?…]+|$)\s*/g) ?? [inner];
+	const parts: string[] = [];
+	let current = "";
+	let budget = firstBudget;
+	for (const sentence of sentences) {
+		const candidate = `${current}${sentence}`;
+		if (current && spokenSeconds(candidate) > budget) {
+			parts.push(current.trim());
+			current = sentence;
+			budget = laterBudget;
+		} else {
+			current = candidate;
+		}
+	}
+	if (current.trim()) parts.push(current.trim());
+	return wrap ? parts.map((p) => `<${wrap[1]}>${p}</${wrap[1]}>`) : parts;
+}
+
+interface Beat {
+	/** the first line is the turn the shot is about; later lines are short reactions */
+	lines: { speaker: string; text: string }[];
+	directions: string[];
+}
+
+function beatSeconds(beat: Beat): number {
+	return estimateSpokenSeconds(beat.lines.map((l) => l.text));
+}
+
+function titleCase(name: string): string {
+	return name.toLowerCase().replace(/(^|[\s'-])(\p{L})/gu, (_m, sep: string, ch: string) => sep + ch.toUpperCase());
+}
+
+function joinNatural(items: string[]): string {
+	return items.length <= 2 ? items.join(" and ") : `${items.slice(0, -1).join(", ")}, and ${items[items.length - 1]}`;
+}
+
+function shotFromBeat(
+	beat: Beat,
+	opts: ImportOptions,
+	totalSeconds: number,
+	continued = false,
+	truncated = false,
+): Shot {
+	const speakers = [...new Set(beat.lines.map((l) => l.speaker))];
+	const subjectIds = [
+		...new Set(speakers.map((s) => opts.speakers[s]?.subject).filter((s): s is string => Boolean(s))),
+	];
+	const named = speakers.map((s) => opts.speakers[s]?.subject ?? titleCase(s));
+
+	let description: string;
+	if (continued) {
+		description = `Continuing the same shot on ${joinNatural(named)}, still speaking.`;
+	} else {
+		description =
+			speakers.length > 1
+				? `${joinNatural(named)} in frame, exchanging lines.`
+				: `Close on ${named[0]}, speaking.`;
+	}
+	if (beat.directions.length) description += ` Stage direction: ${beat.directions.join("; ")}.`;
+
+	const shot = makeShot({
+		type: speakers.length > 1 ? "medium" : "close-up",
+		movement: totalSeconds >= 5 ? "dolly" : "static",
+		subjectIds,
+		description,
+		lines: beat.lines.map((l) => newLine({ speaker: opts.speakers[l.speaker]?.tag ?? "", text: l.text })),
+	});
+	// The guide's marker for speech truncated by the video ending: this clip
+	// stops mid-speech and the next one picks it up.
+	if (truncated) shot.body.push(newTextPart("<cutoff>"));
+	return shot;
+}
+
+/** One shot per speaking turn; a short reaction ("Yeah.") rides along in the shot it reacts to. */
+function groupBeats(turns: ScriptTurn[]): Beat[] {
+	const beats: Beat[] = [];
+	for (const turn of turns) {
+		const { text, directions } = convertDirections(turn.text);
+		if (!text) continue;
+		const last = beats[beats.length - 1];
+		if (last && wordCount(text) <= REACTION_MAX_WORDS) {
+			last.lines.push({ speaker: turn.speaker, text });
+			last.directions.push(...directions);
+		} else {
+			beats.push({ lines: [{ speaker: turn.speaker, text }], directions: [...directions] });
+		}
+	}
+	return beats;
+}
+
+/**
+ * The guide assigns speaker IDs by who first speaks in *each video*, so a
+ * clip that opens on the second character must call them S1. Renumber a
+ * clip's default S1/S2/... tags in order of first appearance; custom tags
+ * (anything not of the form S<n>) are left alone.
+ */
+function renumberSpeakerTags(clip: ImportClip): void {
+	const lines = clip.shots.flatMap((s) => shotLines(s));
+	const seen: string[] = [];
+	for (const l of lines) if (l.speaker && !seen.includes(l.speaker)) seen.push(l.speaker);
+	if (seen.length === 0 || !seen.every((t) => /^S\d+$/.test(t))) return;
+	const renamed = new Map(seen.map((t, i) => [t, `S${i + 1}`]));
+	for (const l of lines) if (l.speaker) l.speaker = renamed.get(l.speaker) ?? l.speaker;
+}
+
+/**
+ * Pack shots into clips that fit `targetSeconds`. Clips break only at cuts
+ * (a cut should bring new information), so a turn is never divided just to
+ * balance lengths. The one exception is a single turn longer than a whole
+ * clip: it is split at sentence breaks, and each later piece starts a
+ * `continuation` clip. The first clip opens with an establishing shot.
+ */
+export function buildClips(turns: ScriptTurn[], opts: ImportOptions): ImportClip[] {
+	const beats = groupBeats(turns);
+
+	const allSubjects = [
+		...new Set(Object.values(opts.speakers).map((s) => s.subject).filter((s): s is string => Boolean(s))),
+	];
+	const establishing = makeShot({
+		type: "establishing",
+		subjectIds: allSubjects,
+		description: `Establishing view of the setting${allSubjects.length ? ` with ${joinNatural(allSubjects)} in frame` : ""}.`,
+	});
+
+	const clips: ImportClip[] = [];
+	let current: ImportClip = { shots: [establishing], continuation: false };
+	let seconds = ESTABLISHING_SECONDS;
+	const hasDialogue = () => current.shots.some((s) => shotLines(s).length > 0);
+	const close = (continuation: boolean) => {
+		clips.push(current);
+		current = { shots: [], continuation };
+		seconds = 0;
+	};
+
+	for (const beat of beats) {
+		const total = beatSeconds(beat);
+
+		if (total <= opts.targetSeconds) {
+			if (hasDialogue() && seconds + total > opts.targetSeconds) close(false);
+			current.shots.push(shotFromBeat(beat, opts, total));
+			seconds += total;
+			continue;
+		}
+
+		// Longer than any clip can hold: the split has to fall inside this shot.
+		if (current.shots.length > 0 && opts.targetSeconds - seconds < MIN_PART_SECONDS) close(false);
+		const [main, ...reactions] = beat.lines;
+		const parts = splitToBudgets(main.text, opts.targetSeconds - seconds, opts.targetSeconds);
+		parts.forEach((part, k) => {
+			if (k > 0) close(true);
+			const lines = [{ ...main, text: part }, ...(k === parts.length - 1 ? reactions : [])];
+			const pieceBeat: Beat = { lines, directions: k === 0 ? beat.directions : [] };
+			current.shots.push(shotFromBeat(pieceBeat, opts, total, k > 0, k < parts.length - 1));
+			seconds += beatSeconds(pieceBeat);
+		});
+	}
+
+	if (current.shots.length > 0 && (clips.length === 0 || hasDialogue())) clips.push(current);
+	for (const clip of clips) renumberSpeakerTags(clip);
+	return clips;
+}
diff --git a/apps/fastvideo_studio/src/lib/h3Shots.test.ts b/apps/fastvideo_studio/src/lib/h3Shots.test.ts
new file mode 100644
index 000000000..473718089
--- /dev/null
+++ b/apps/fastvideo_studio/src/lib/h3Shots.test.ts
@@ -0,0 +1,336 @@
+import { describe, expect, it } from "vitest";
+import {
+	detailedDescriptionFromShots,
+	formatCutTime,
+	makeShot,
+	newLine,
+	parseCutTime,
+	parseDetailedDescription,
+	parseSubjectLabels,
+	resolveShotStarts,
+	retentionAnalysisFromShots,
+	shotLines,
+	shotProse,
+	shotsFromDetailedDescription,
+	type Shot,
+	type ShotInit,
+} from "@/lib/h3Shots";
+
+function shot(init: ShotInit): Shot {
+	return makeShot(init);
+}
+
+describe("parseSubjectLabels", () => {
+	it("returns ordered, de-duplicated <Subject N> labels", () => {
+		const text = "<Subject 1> is the dog.\n<Subject 2> is the cat. <Subject 1> again.";
+		expect(parseSubjectLabels(text)).toEqual(["<Subject 1>", "<Subject 2>"]);
+	});
+
+	it("returns an empty list when there are none", () => {
+		expect(parseSubjectLabels("no subjects here")).toEqual([]);
+	});
+});
+
+describe("detailedDescriptionFromShots", () => {
+	it("serializes shot type, movement, description, and dialogue", () => {
+		const shots = [
+			shot({
+				type: "close-up",
+				movement: "handheld",
+				description: "She looks up.",
+				lines: [newLine({ speaker: "S1", text: "I found it." })],
+			}),
+		];
+		expect(detailedDescriptionFromShots(shots)).toBe(
+			"[Shot 1] Close-up shot, handheld. She looks up. (S1) <d>[English] I found it.</d>",
+		);
+	});
+
+	it("omits the movement clause for a static camera", () => {
+		const shots = [shot({ type: "wide", movement: "static", description: "An empty street." })];
+		expect(detailedDescriptionFromShots(shots)).toBe("[Shot 1] Wide shot. An empty street.");
+	});
+
+	it("numbers shots in order, one per line, and marks later shots with a cut time", () => {
+		const shots = [shot({ description: "First." }), shot({ description: "Second." })];
+		const out = detailedDescriptionFromShots(shots);
+		expect(out).toBe(
+			"[Shot 1] Medium shot. First.\n[Shot 2] At 00:02.000, the shot cuts to a medium shot. Second.",
+		);
+	});
+
+	it("uses the right article and movement in the cut wording", () => {
+		const shots = [
+			shot({ type: "wide" }),
+			shot({ type: "establishing", startSeconds: 3 }),
+			shot({ type: "extreme-close-up", movement: "dolly", startSeconds: 5.25 }),
+			shot({ type: "pov", startSeconds: 7 }),
+		];
+		const out = detailedDescriptionFromShots(shots).split("\n");
+		expect(out[1]).toBe("[Shot 2] At 00:03.000, the shot cuts to an establishing shot.");
+		expect(out[2]).toBe("[Shot 3] At 00:05.250, the shot cuts to an extreme close-up shot, dolly.");
+		expect(out[3]).toBe("[Shot 4] At 00:07.000, the shot cuts to a POV shot.");
+	});
+
+	it("writes the preamble before the first shot", () => {
+		const out = detailedDescriptionFromShots([shot({ description: "A." })], "Keep the film-noir look.");
+		expect(out).toBe("Keep the film-noir look.\n[Shot 1] Medium shot. A.");
+	});
+});
+
+describe("cut times", () => {
+	it("formats and parses MM:SS.mmm", () => {
+		expect(formatCutTime(3.5)).toBe("00:03.500");
+		expect(formatCutTime(75.04)).toBe("01:15.040");
+		expect(parseCutTime("00:03.500")).toBe(3.5);
+		expect(parseCutTime("0:03.5")).toBe(3.5);
+		expect(parseCutTime("3.5")).toBe(3.5);
+		expect(parseCutTime("1:15")).toBe(75);
+		expect(parseCutTime("soon")).toBeNull();
+	});
+
+	it("starts the opening shot at 0 and follows estimated shot lengths", () => {
+		const shots = [
+			shot({ lines: [newLine({ text: "A short line." })] }),
+			shot({}),
+			shot({}),
+		];
+		const starts = resolveShotStarts(shots);
+		expect(starts[0]).toBe(0);
+		expect(starts[1]).toBeGreaterThanOrEqual(2);
+		expect(starts[2]).toBeCloseTo(starts[1] + 2, 3);
+	});
+
+	it("honors an explicit time and keeps times strictly increasing", () => {
+		const shots = [shot({}), shot({ startSeconds: 4 }), shot({ startSeconds: 1 })];
+		const starts = resolveShotStarts(shots);
+		expect(starts[1]).toBe(4);
+		expect(starts[2]).toBeGreaterThan(starts[1]);
+	});
+});
+
+describe("shotsFromDetailedDescription", () => {
+	it("round-trips type, movement, dialogue, speaker, and subjects", () => {
+		const original = [
+			shot({
+				type: "extreme-close-up",
+				movement: "dolly",
+				description: "Her hand trembles.",
+				lines: [newLine({ speaker: "S1", text: "Wait." })],
+				subjectIds: ["<Subject 1>"],
+			}),
+		];
+		const serialized = detailedDescriptionFromShots(original);
+		const withSubject = serialized.replace("Her hand trembles.", "Her hand trembles, <Subject 1> hesitates.");
+		const parsed = shotsFromDetailedDescription(withSubject, ["<Subject 1>", "<Subject 2>"]);
+
+		expect(parsed).toHaveLength(1);
+		expect(parsed[0].type).toBe("extreme-close-up");
+		expect(parsed[0].movement).toBe("dolly");
+		expect(shotLines(parsed[0])).toHaveLength(1);
+		expect(shotLines(parsed[0])[0].speaker).toBe("S1");
+		expect(shotLines(parsed[0])[0].text).toBe("Wait.");
+		expect(parsed[0].subjectIds).toEqual(["<Subject 1>"]);
+	});
+
+	it("keeps every dialogue line of an exchange, in order", () => {
+		const original = [
+			shot({
+				description: "They talk.",
+				lines: [
+					newLine({ speaker: "S1", text: "Ready?" }),
+					newLine({ speaker: "S2", text: "<pause> Always." }),
+					newLine({ speaker: "S1", text: "Then go." }),
+				],
+			}),
+		];
+		const serialized = detailedDescriptionFromShots(original);
+		expect(serialized).toBe(
+			"[Shot 1] Medium shot. They talk. (S1) <d>[English] Ready?</d> (S2) <d>[English] <pause> Always.</d> (S1) <d>[English] Then go.</d>",
+		);
+
+		const parsed = shotsFromDetailedDescription(serialized, []);
+		expect(shotProse(parsed[0])).toBe("They talk.");
+		expect(shotLines(parsed[0]).map((l) => [l.speaker, l.text])).toEqual([
+			["S1", "Ready?"],
+			["S2", "<pause> Always."],
+			["S1", "Then go."],
+		]);
+	});
+
+	it("round-trips a non-English language marker instead of relabelling it English", () => {
+		const original = [shot({ lines: [newLine({ speaker: "S1", text: "你好。", language: "Chinese" })] })];
+		const serialized = detailedDescriptionFromShots(original);
+		expect(serialized).toContain("<d>[Chinese] 你好。</d>");
+		expect(shotLines(shotsFromDetailedDescription(serialized, [])[0])[0].language).toBe("Chinese");
+	});
+
+	it("round-trips a multi-shot list without pinning the automatic cut times", () => {
+		const original = [
+			shot({ type: "establishing", description: "A rooftop." }),
+			shot({
+				type: "close-up",
+				movement: "handheld",
+				description: "He turns.",
+				lines: [newLine({ speaker: "S2", text: "I had no choice." })],
+			}),
+			shot({ type: "wide", movement: "dolly", description: "They stand apart." }),
+		];
+		const serialized = detailedDescriptionFromShots(original);
+		const parsed = shotsFromDetailedDescription(serialized, []);
+
+		expect(parsed.map((p) => [p.type, p.movement, p.freeLead])).toEqual([
+			["establishing", "static", false],
+			["close-up", "handheld", false],
+			["wide", "dolly", false],
+		]);
+		expect(parsed.every((p) => p.startSeconds === null)).toBe(true);
+		expect(detailedDescriptionFromShots(parsed)).toBe(serialized);
+	});
+
+	it("pins a cut time that differs from the automatic one", () => {
+		const parsed = shotsFromDetailedDescription(
+			"[Shot 1] Wide shot. A dock.\n[Shot 2] At 00:07.400, the shot cuts to a close-up shot. A face.",
+			[],
+		);
+		expect(parsed[1].startSeconds).toBe(7.4);
+	});
+
+	it("keeps hand-written shots verbatim, with the preamble and cut times", () => {
+		const handWritten = [
+			"The target video keeps the live-action style of <Video 1> unchanged.",
+			"[Shot 1] A handheld medium shot holds <Subject 1> at frame left, half-occluded by plastic sheeting.",
+			"[Shot 2] At 00:03.800, the shot cuts to a low tracking shot behind the boots of <Subject 1> as he sprints.",
+			"[Shot 3] At 00:07.400, the camera cuts to a close front-on tracking shot. He (S1) says: <d>[English] Go.</d> <cutoff>",
+		].join("\n");
+		const { preamble, shots } = parseDetailedDescription(handWritten, ["<Subject 1>"]);
+
+		expect(preamble).toBe("The target video keeps the live-action style of <Video 1> unchanged.");
+		expect(shots).toHaveLength(3);
+		expect(shots.every((s) => s.freeLead)).toBe(true);
+		expect(shots.map((s) => s.startSeconds)).toEqual([null, 3.8, 7.4]);
+		expect(shots[0].subjectIds).toEqual(["<Subject 1>"]);
+		// Dialogue in a hand-written shot is pulled out into an editable line...
+		expect(shotLines(shots[2]).map((l) => [l.speaker, l.text])).toEqual([["", "Go."]]);
+		expect(detailedDescriptionFromShots(shots, preamble)).toBe(handWritten);
+	});
+
+	it("keeps a trailing tag after the dialogue it modifies", () => {
+		const text =
+			"[Shot 1] He (S1) says: <d>[English] <breath> I think you're... <i>sorry</i>.</d> <cutoff>";
+		const [shot1] = shotsFromDetailedDescription(text, []);
+		expect(shot1.freeLead).toBe(true);
+		expect(shot1.body.map((p) => p.kind)).toEqual(["text", "line", "text"]);
+		expect(shotLines(shot1)[0].text).toBe("<breath> I think you're... <i>sorry</i>.");
+		expect(detailedDescriptionFromShots([shot1])).toBe(text);
+	});
+
+	it("keeps interleaved dialogue in place, whitespace and all, in hand-written shots", () => {
+		const text = "[Shot 1] She waits.\n  (S1) <d>[English] Ready?</d>\nHe nods.  (S2) <d>[English] Ready.</d>  The door opens.";
+		const [s1] = shotsFromDetailedDescription(text, []);
+		expect(s1.body.map((p) => p.kind)).toEqual(["text", "line", "text", "line", "text"]);
+		expect(detailedDescriptionFromShots([s1])).toBe(text);
+	});
+
+	it("keeps dialogue between prose in shots this editor wrote too", () => {
+		const text = "[Shot 1] Wide shot. He looks up. (S1) <d>[English] Hello.</d> Then he leaves.";
+		const [s1] = shotsFromDetailedDescription(text, []);
+		expect(s1.freeLead).toBe(false);
+		expect(s1.body.map((p) => p.kind)).toEqual(["text", "line", "text"]);
+		expect(detailedDescriptionFromShots([s1])).toBe(text);
+	});
+
+	it("edits a line without moving the text or lines around it", () => {
+		const text = "[Shot 1] Wide shot. A. (S1) <d>[English] One.</d> B. (S2) <d>[English] Two.</d> C.";
+		const [s1] = shotsFromDetailedDescription(text, []);
+		const second = shotLines(s1)[1];
+		second.text = "Changed.";
+		expect(detailedDescriptionFromShots([s1])).toBe(
+			"[Shot 1] Wide shot. A. (S1) <d>[English] One.</d> B. (S2) <d>[English] Changed.</d> C.",
+		);
+	});
+
+	it("puts a newly added line at the end and separates it from the prose", () => {
+		const text = "[Shot 1] Wide shot. He speaks. (S1) <d>[English] First.</d> He pauses.";
+		const [s1] = shotsFromDetailedDescription(text, []);
+		s1.body.push({ kind: "line", line: newLine({ speaker: "S1", text: "Later." }) });
+		expect(detailedDescriptionFromShots([s1])).toBe(
+			"[Shot 1] Wide shot. He speaks. (S1) <d>[English] First.</d> He pauses. (S1) <d>[English] Later.</d>",
+		);
+	});
+
+	it("reordering blocks changes the written order", () => {
+		const [s1] = shotsFromDetailedDescription(
+			"[Shot 1] Wide shot. (S1) <d>[English] One.</d> (S2) <d>[English] Two.</d>",
+			[],
+		);
+		const lineBlocks = s1.body.filter((p) => p.kind === "line");
+		s1.body = [...s1.body.filter((p) => p.kind !== "line"), lineBlocks[1], lineBlocks[0]];
+		expect(detailedDescriptionFromShots([s1])).toBe(
+			"[Shot 1] Wide shot. (S2) <d>[English] Two.</d> (S1) <d>[English] One.</d>",
+		);
+	});
+
+	it("recovers type and movement from the guide's cut wording", () => {
+		const parsed = shotsFromDetailedDescription(
+			"[Shot 1] Wide shot.\n[Shot 2] At 00:04.000, the camera cuts to an over-the-shoulder shot, handheld. Two people talk.",
+			[],
+		);
+		expect(parsed[1].type).toBe("over-the-shoulder");
+		expect(parsed[1].movement).toBe("handheld");
+		expect(shotProse(parsed[1])).toBe("Two people talk.");
+		expect(parsed[1].freeLead).toBe(false);
+	});
+
+	it("falls back to defaults for arbitrary prose, keeping it as the description", () => {
+		const parsed = shotsFromDetailedDescription("A toy car drives into a plush dog.", []);
+		expect(parsed).toHaveLength(1);
+		expect(parsed[0].type).toBe("medium");
+		expect(parsed[0].movement).toBe("static");
+		expect(shotProse(parsed[0])).toBe("A toy car drives into a plush dog.");
+	});
+
+	it("returns an empty list for blank text", () => {
+		expect(shotsFromDetailedDescription("", [])).toEqual([]);
+	});
+});
+
+describe("retentionAnalysisFromShots", () => {
+	it("generates an appears-in line per subject with shots", () => {
+		const shots = [
+			shot({ subjectIds: ["<Subject 1>"] }),
+			shot({ subjectIds: ["<Subject 1>", "<Subject 2>"] }),
+		];
+		const out = retentionAnalysisFromShots(shots, ["<Subject 1>", "<Subject 2>"], "");
+		expect(out).toBe(
+			[
+				"<Subject 1> (appears in [Shot 1], [Shot 2]): fully_preserved - what is retained.",
+				"<Subject 2> (appears in [Shot 2]): fully_preserved - what is retained.",
+			].join("\n"),
+		);
+	});
+
+	it("preserves an existing retention note instead of overwriting it", () => {
+		const shots = [shot({ subjectIds: ["<Subject 1>"] })];
+		const previous = "<Subject 1> (appears in [Shot 1]): fully_preserved - fur and collar retained.";
+		const out = retentionAnalysisFromShots(shots, ["<Subject 1>"], previous);
+		expect(out).toBe("<Subject 1> (appears in [Shot 1]): fully_preserved - fur and collar retained.");
+	});
+
+	it("passes through lines it doesn't own, like audio references", () => {
+		const shots = [shot({ subjectIds: ["<Subject 1>"] })];
+		const previous = [
+			"<Subject 1> (appears in [Shot 3]): fully_preserved - old note.",
+			"<Audio 1>: reference - how it guides the audio.",
+		].join("\n");
+		const out = retentionAnalysisFromShots(shots, ["<Subject 1>"], previous);
+		expect(out).toContain("<Subject 1> (appears in [Shot 1]): fully_preserved - old note.");
+		expect(out).toContain("<Audio 1>: reference - how it guides the audio.");
+	});
+
+	it("omits a subject with no shots rather than emitting an empty appears-in", () => {
+		const shots = [shot({ subjectIds: [] })];
+		const out = retentionAnalysisFromShots(shots, ["<Subject 1>"], "");
+		expect(out).toBe("");
+	});
+});
diff --git a/apps/fastvideo_studio/src/lib/h3Shots.ts b/apps/fastvideo_studio/src/lib/h3Shots.ts
new file mode 100644
index 000000000..c159fe9f2
--- /dev/null
+++ b/apps/fastvideo_studio/src/lib/h3Shots.ts
@@ -0,0 +1,478 @@
+/**
+ * MiniMax-H3 shot list: a structured camera-direction editor that serializes
+ * into `detailed_description`'s `[Shot N]` blocks and keeps
+ * `retention_analysis`'s "appears in [...]" lines in sync, per the model's
+ * reference prompt guide (see h3Prompt.ts / h3References.ts for the six
+ * guided sections this plugs into).
+ */
+
+import { SPEECH_RATES, estimateDuration } from "@/lib/h3DialogueEstimate";
+
+export type ShotType =
+	| "establishing"
+	| "wide"
+	| "medium"
+	| "close-up"
+	| "extreme-close-up"
+	| "over-the-shoulder"
+	| "pov"
+	| "insert";
+
+export type CameraMovement =
+	| "static"
+	| "pan"
+	| "tilt"
+	| "dolly"
+	| "zoom"
+	| "handheld"
+	| "tracking"
+	| "crane";
+
+export interface DialogueLine {
+	/** stable key for React lists -- not part of the serialized prompt */
+	id: string;
+	/** speaker tag without parens, e.g. "S1" -- blank omits the "(S1)" prefix */
+	speaker: string;
+	/** spoken text only (speech tags allowed); wrapped in <d>[language] ...</d> on serialize */
+	text: string;
+	/** language marker written inside <d>; recovered on parse so non-English dialogue round-trips */
+	language: string;
+}
+
+/** A run of prose inside a shot (composition, action, environment, delivery notes). */
+export interface TextPart {
+	kind: "text";
+	/** stable key for React lists -- not part of the serialized prompt */
+	id: string;
+	text: string;
+}
+
+/** One spoken line, written as `(S1) <d>[English] ...</d>`. Its position among the parts is where it plays. */
+export interface LinePart {
+	kind: "line";
+	line: DialogueLine;
+}
+
+export type ShotPart = TextPart | LinePart;
+
+export interface Shot {
+	/** stable key for React lists -- not part of the serialized prompt */
+	id: string;
+	type: ShotType;
+	movement: CameraMovement;
+	/** <Subject N> labels (from subject_definitions) that appear in this shot */
+	subjectIds: string[];
+	/**
+	 * The shot's content in the order it reads: prose and spoken lines
+	 * interleaved. This one list is the only record of where each line falls,
+	 * so a `<d>` block stays mid-sentence if that is where it was written.
+	 */
+	body: ShotPart[];
+	/** cut-in time in seconds; null = computed from the preceding shots. Ignored for the opening shot. */
+	startSeconds: number | null;
+	/**
+	 * The description carries its own lead-in (text opened from a hand-written
+	 * prompt) so type/movement are not written and the wording is kept verbatim.
+	 */
+	freeLead: boolean;
+}
+
+export const SHOT_TYPES: { id: ShotType; label: string }[] = [
+	{ id: "establishing", label: "Establishing" },
+	{ id: "wide", label: "Wide" },
+	{ id: "medium", label: "Medium" },
+	{ id: "close-up", label: "Close-up" },
+	{ id: "extreme-close-up", label: "Extreme close-up" },
+	{ id: "over-the-shoulder", label: "Over-the-shoulder" },
+	{ id: "pov", label: "POV" },
+	{ id: "insert", label: "Insert" },
+];
+
+export const CAMERA_MOVEMENTS: { id: CameraMovement; label: string }[] = [
+	{ id: "static", label: "Static" },
+	{ id: "pan", label: "Pan" },
+	{ id: "tilt", label: "Tilt" },
+	{ id: "dolly", label: "Dolly" },
+	{ id: "zoom", label: "Zoom" },
+	{ id: "handheld", label: "Handheld" },
+	{ id: "tracking", label: "Tracking" },
+	{ id: "crane", label: "Crane" },
+];
+
+const SHOT_TYPE_LABEL = Object.fromEntries(SHOT_TYPES.map((t) => [t.id, t.label])) as Record<ShotType, string>;
+const MOVEMENT_LABEL = Object.fromEntries(CAMERA_MOVEMENTS.map((m) => [m.id, m.label])) as Record<
+	CameraMovement,
+	string
+>;
+const SHOT_TYPE_BY_LABEL = Object.fromEntries(SHOT_TYPES.map((t) => [t.label.toLowerCase(), t.id])) as Record<
+	string,
+	ShotType
+>;
+const MOVEMENT_BY_LABEL = Object.fromEntries(
+	CAMERA_MOVEMENTS.map((m) => [m.label.toLowerCase(), m.id]),
+) as Record<string, CameraMovement>;
+
+let shotSeq = 0;
+export function newShot(): Shot {
+	shotSeq += 1;
+	return {
+		id: `shot-${Date.now()}-${shotSeq}`,
+		type: "medium",
+		movement: "static",
+		subjectIds: [],
+		body: [newTextPart()],
+		startSeconds: null,
+		freeLead: false,
+	};
+}
+
+let partSeq = 0;
+export function newTextPart(text = ""): TextPart {
+	partSeq += 1;
+	return { kind: "text", id: `text-${Date.now()}-${partSeq}`, text };
+}
+
+export function linePart(line: DialogueLine): LinePart {
+	return { kind: "line", line };
+}
+
+/** Every spoken line of the shot, in playback order. */
+export function shotLines(shot: Shot): DialogueLine[] {
+	return shot.body.flatMap((p) => (p.kind === "line" ? [p.line] : []));
+}
+
+/** The shot's prose with the dialogue left out. */
+export function shotProse(shot: Shot): string {
+	return shot.body
+		.flatMap((p) => (p.kind === "text" ? [p.text.trim()] : []))
+		.filter(Boolean)
+		.join(" ");
+}
+
+/** A body of one prose part followed by the given lines -- the common "description, then dialogue" shape. */
+export function makeBody(description = "", lines: DialogueLine[] = []): ShotPart[] {
+	return [newTextPart(description), ...lines.map(linePart)];
+}
+
+export interface ShotInit extends Partial<Omit<Shot, "id" | "body">> {
+	description?: string;
+	lines?: DialogueLine[];
+	body?: ShotPart[];
+}
+
+/** Build a shot from a description and lines (or an explicit body). */
+export function makeShot(init: ShotInit = {}): Shot {
+	const { description, lines, body, ...rest } = init;
+	return { ...newShot(), ...rest, body: body ?? makeBody(description ?? "", lines ?? []) };
+}
+
+let lineSeq = 0;
+export function newLine(patch: Partial<Omit<DialogueLine, "id">> = {}): DialogueLine {
+	lineSeq += 1;
+	return { id: `line-${Date.now()}-${lineSeq}`, speaker: "", text: "", language: "English", ...patch };
+}
+
+/** Copy shots with brand-new ids so components keyed by id (which own their own text state) remount instead of showing stale edits. */
+export function withFreshIds(shots: Shot[]): Shot[] {
+	return shots.map((s) => ({
+		...s,
+		id: newShot().id,
+		subjectIds: [...s.subjectIds],
+		body: s.body.map((p) =>
+			p.kind === "text" ? { ...p, id: newTextPart().id } : { kind: "line" as const, line: { ...p.line, id: newLine().id } },
+		),
+	}));
+}
+
+/** Ordered, de-duplicated <Subject N> labels mentioned in subject_definitions. */
+export function parseSubjectLabels(subjectDefinitions: string): string[] {
+	const seen: string[] = [];
+	const re = /<Subject\s+\d+>/g;
+	let m: RegExpExecArray | null;
+	while ((m = re.exec(subjectDefinitions))) {
+		if (!seen.includes(m[0])) seen.push(m[0]);
+	}
+	return seen;
+}
+
+/* ---------------------------------------------------------------- */
+/*  Cut times                                                         */
+/* ---------------------------------------------------------------- */
+
+const MIN_SHOT_SECONDS = 2;
+const LINE_GAP_SECONDS = 0.5;
+
+/** Ballpark seconds a shot with these spoken lines runs; drives auto cut times. */
+export function estimateSpokenSeconds(texts: string[]): number {
+	const spoken = texts.filter((t) => t.trim());
+	const speech = spoken.reduce((n, t) => n + estimateDuration(t, SPEECH_RATES.Natural, 1).total, 0);
+	return Math.max(MIN_SHOT_SECONDS, speech + (spoken.length ? LINE_GAP_SECONDS : 0));
+}
+
+export function estimateShotSeconds(shot: Shot): number {
+	return estimateSpokenSeconds(shotLines(shot).map((l) => l.text));
+}
+
+/**
+ * Cut-in time of every shot: the opening shot is 0, later ones use their
+ * explicit time or follow the previous shot's estimated length. Times are
+ * kept strictly increasing, as the guide requires.
+ */
+export function resolveShotStarts(shots: Shot[]): number[] {
+	const starts: number[] = [];
+	shots.forEach((shot, i) => {
+		if (i === 0) {
+			starts.push(0);
+			return;
+		}
+		const prev = starts[i - 1];
+		let start = shot.startSeconds ?? prev + estimateShotSeconds(shots[i - 1]);
+		if (start <= prev) start = prev + 0.1;
+		starts.push(Math.round(start * 1000) / 1000);
+	});
+	return starts;
+}
+
+/** MM:SS.mmm, the format the guide uses for cut times ("At 00:03.500, ..."). */
+export function formatCutTime(seconds: number): string {
+	const totalMs = Math.max(0, Math.round(seconds * 1000));
+	const pad = (n: number, w: number) => String(n).padStart(w, "0");
+	const totalSec = Math.floor(totalMs / 1000);
+	return `${pad(Math.floor(totalSec / 60), 2)}:${pad(totalSec % 60, 2)}.${pad(totalMs % 1000, 3)}`;
+}
+
+/** Accepts "3.5", "0:03.5", or "00:03.500"; returns seconds, or null if unreadable. */
+export function parseCutTime(text: string): number | null {
+	const m = text.trim().match(/^(?:(\d+):)?(\d+(?:\.\d+)?)$/);
+	if (!m) return null;
+	return (m[1] ? Number(m[1]) * 60 : 0) + Number(m[2]);
+}
+
+/* ---------------------------------------------------------------- */
+/*  Serialize                                                         */
+/* ---------------------------------------------------------------- */
+
+/** `raw` keeps a line's text exactly as stored (hand-written shots); otherwise it is trimmed. */
+function partText(part: ShotPart, raw = false): string {
+	if (part.kind === "text") return part.text;
+	const { speaker, language, text } = part.line;
+	if (!text.trim()) return "";
+	return `${speaker.trim() ? `(${speaker.trim()}) ` : ""}<d>[${language.trim() || "English"}] ${raw ? text : text.trim()}</d>`;
+}
+
+/** Concatenate pieces, adding one space only where neither side already has whitespace. */
+function joinPieces(pieces: string[]): string {
+	let out = "";
+	for (const piece of pieces) {
+		if (!piece) continue;
+		if (out && !/\s$/.test(out) && !/^\s/.test(piece)) out += " ";
+		out += piece;
+	}
+	return out;
+}
+
+function serializeShot(shot: Shot, n: number, start: number): string {
+	// [Shot 1] has no timestamp; later shots mark their cut time.
+	const cut = n > 1 ? `At ${formatCutTime(start)}, ` : "";
+
+	if (shot.freeLead) {
+		// Text parts keep their exact whitespace, so hand-written wording round-trips byte for byte.
+		return `[Shot ${n}] ${cut}${joinPieces(shot.body.map((p) => partText(p, true)))}`.trim();
+	}
+
+	const body = shot.body
+		.map((p) => partText(p).trim())
+		.filter(Boolean)
+		.join(" ");
+	const label = SHOT_TYPE_LABEL[shot.type];
+	const movementPart = shot.movement === "static" ? "" : `, ${MOVEMENT_LABEL[shot.movement].toLowerCase()}`;
+	const lead =
+		n > 1
+			? `the shot cuts to ${/^[aeiou]/i.test(label) ? "an" : "a"} ${label === "POV" ? label : label.toLowerCase()} shot${movementPart}.`
+			: `${label} shot${movementPart}.`;
+	return [`[Shot ${n}] ${cut}${lead}`, body].filter(Boolean).join(" ");
+}
+
+/**
+ * Serialize the shot list into detailed_description's [Shot N] format:
+ * `[Shot 1] ...` then `[Shot N] At MM:SS.mmm, the shot cuts to ...`. Optional
+ * `preamble` (style notes that come before the first shot) is written first.
+ */
+export function detailedDescriptionFromShots(shots: Shot[], preamble = ""): string {
+	const starts = resolveShotStarts(shots);
+	return [preamble.trim(), ...shots.map((s, i) => serializeShot(s, i + 1, starts[i]))]
+		.filter(Boolean)
+		.join("\n");
+}
+
+/* ---------------------------------------------------------------- */
+/*  Parse                                                             */
+/* ---------------------------------------------------------------- */
+
+const CUT_TIME_RE = /^At (\d+):(\d{2})(?:\.(\d{1,3}))?,\s*/i;
+// "[the shot cuts to] [a] <type> shot[, <movement>]." -- the lead-in this
+// editor writes, recovered on parse so type/movement survive a reopen.
+const LEAD_RE =
+	/^(?:(?:the (?:shot|camera) (?:cuts|transitions|changes|switches) to |cuts to )?(?:an? )?)([a-zA-Z -]+?) shot(?:,\s*([a-zA-Z]+))?\.\s*/i;
+const DIALOGUE_RE = /(?:\(([^)]+)\)\s*)?<d>(?:\[([^\]]*)\]\s*)?([\s\S]*?)<\/d>/g;
+
+/**
+ * Split shot text at its `<d>` blocks into an ordered body. `exact` keeps the
+ * surrounding text untouched (whitespace included) so it can be rebuilt
+ * byte for byte; otherwise prose is trimmed and empty runs dropped.
+ */
+function splitBody(text: string, exact: boolean): ShotPart[] {
+	const parts: ShotPart[] = [];
+	const pushText = (raw: string) => {
+		const t = exact ? raw : raw.trim();
+		if (t) parts.push(newTextPart(t));
+	};
+	let last = 0;
+	for (const m of text.matchAll(DIALOGUE_RE)) {
+		const start = m.index ?? 0;
+		pushText(text.slice(last, start));
+		parts.push(
+			linePart(
+				newLine({
+					speaker: m[1]?.trim() ?? "",
+					language: m[2]?.trim() || "English",
+					text: exact ? (m[3] ?? "") : (m[3] ?? "").trim(),
+				}),
+			),
+		);
+		last = start + m[0].length;
+	}
+	pushText(text.slice(last));
+	return parts;
+}
+
+function parseShotBlock(block: string, subjectLabels: string[]): { shot: Shot; time: number | null } {
+	const shot = newShot();
+	shot.subjectIds = subjectLabels.filter((label) => block.includes(label));
+	let rest = block;
+
+	let time: number | null = null;
+	const t = rest.match(CUT_TIME_RE);
+	if (t) {
+		time = Number(t[1]) * 60 + Number(t[2]) + (t[3] ? Number(`0.${t[3]}`) : 0);
+		rest = rest.slice(t[0].length);
+	}
+
+	const lead = rest.match(LEAD_RE);
+	const typeId = lead ? SHOT_TYPE_BY_LABEL[lead[1].trim().toLowerCase()] : undefined;
+	if (lead && typeId) {
+		shot.type = typeId;
+		const moveLabel = lead[2]?.trim().toLowerCase();
+		if (moveLabel && MOVEMENT_BY_LABEL[moveLabel]) shot.movement = MOVEMENT_BY_LABEL[moveLabel];
+		rest = rest.slice(lead[0].length);
+		shot.body = splitBody(rest, false);
+	} else {
+		// Hand-written wording: keep it exactly as written. Dialogue is still
+		// pulled out into editable lines, but each stays where it was written.
+		shot.freeLead = true;
+		shot.body = splitBody(rest, true);
+	}
+	if (shot.body.length === 0) shot.body = [newTextPart()];
+	return { shot, time };
+}
+
+/**
+ * Split an existing detailed_description back into shots so opening the
+ * builder on a job someone already wrote doesn't discard it. Shots are found
+ * by their line-start [Shot N] markers; anything before the first is kept as
+ * the preamble. Only wording this editor generated is decomposed into
+ * type/movement/dialogue -- anything else is kept verbatim.
+ */
+export function parseDetailedDescription(
+	text: string,
+	subjectLabels: string[],
+): { preamble: string; shots: Shot[] } {
+	const markers = [...text.matchAll(/^\[Shot\s+\d+\][ \t]*/gm)];
+	if (markers.length === 0) {
+		const body = text.trim();
+		if (!body) return { preamble: "", shots: [] };
+		const { shot } = parseShotBlock(body, subjectLabels);
+		shot.freeLead = true;
+		shot.body = splitBody(body, true);
+		return { preamble: "", shots: [shot] };
+	}
+
+	const preamble = text.slice(0, markers[0].index).trim();
+	const parsed = markers.map((m, i) => {
+		const from = (m.index ?? 0) + m[0].length;
+		const to = i + 1 < markers.length ? (markers[i + 1].index ?? text.length) : text.length;
+		return parseShotBlock(text.slice(from, to).trim(), subjectLabels);
+	});
+	const shots = parsed.map((p) => p.shot);
+
+	// Only pin a cut time when it differs from what the editor would compute,
+	// so text this editor wrote keeps flowing as earlier shots change.
+	const starts: number[] = [0];
+	shots.forEach((shot, i) => {
+		if (i === 0) return;
+		const auto = starts[i - 1] + estimateShotSeconds(shots[i - 1]);
+		const time = parsed[i].time;
+		if (time == null) {
+			starts.push(auto);
+			return;
+		}
+		starts.push(time);
+		if (Math.abs(time - auto) > 0.001) shot.startSeconds = time;
+	});
+	return { preamble, shots };
+}
+
+export function shotsFromDetailedDescription(text: string, subjectLabels: string[]): Shot[] {
+	return parseDetailedDescription(text, subjectLabels).shots;
+}
+
+const RETENTION_LINE_RE = /^(<[^>]+>)\s*\(appears in [^)]*\):\s*(.*)$/gm;
+const DEFAULT_RETENTION_NOTE = "fully_preserved - what is retained.";
+
+/** label -> free-text retention note (everything after the colon), for lines this parser recognizes. */
+function parseRetentionNotes(retentionAnalysis: string): Map<string, string> {
+	const notes = new Map<string, string>();
+	let m: RegExpExecArray | null;
+	RETENTION_LINE_RE.lastIndex = 0;
+	while ((m = RETENTION_LINE_RE.exec(retentionAnalysis))) notes.set(m[1], m[2].trim());
+	return notes;
+}
+
+/**
+ * Rebuild retention_analysis's shot-tracked lines from the shot list. Each
+ * subject's existing retention note (the free text after the colon) is kept
+ * untouched -- only the "(appears in [...])" shot list is regenerated, using
+ * a sensible default note for a subject that's newly shot-tracked. Anything
+ * this shot list doesn't own (untracked subjects, audio-reference lines,
+ * free text) passes through unchanged rather than being discarded.
+ */
+export function retentionAnalysisFromShots(
+	shots: Shot[],
+	subjectLabels: string[],
+	previousRetentionAnalysis: string,
+): string {
+	const existingNotes = parseRetentionNotes(previousRetentionAnalysis);
+	const regenerated = new Set<string>();
+	const shotLines: string[] = [];
+
+	for (const label of subjectLabels) {
+		const shotNums = shots
+			.map((s, i) => (s.subjectIds.includes(label) ? i + 1 : null))
+			.filter((n): n is number => n !== null);
+		if (shotNums.length === 0) continue;
+		regenerated.add(label);
+		const note = existingNotes.get(label) ?? DEFAULT_RETENTION_NOTE;
+		shotLines.push(`${label} (appears in ${shotNums.map((n) => `[Shot ${n}]`).join(", ")}): ${note}`);
+	}
+
+	const passthroughLines = previousRetentionAnalysis
+		.split("\n")
+		.map((l) => l.trim())
+		.filter((line) => {
+			if (!line) return false;
+			const m = line.match(/^(<[^>]+>)\s*\(appears in/);
+			return !(m && regenerated.has(m[1]));
+		});
+
+	return [...shotLines, ...passthroughLines].join("\n");
+}
diff --git a/apps/fastvideo_studio/src/lib/h3Tags.test.ts b/apps/fastvideo_studio/src/lib/h3Tags.test.ts
new file mode 100644
index 000000000..72bfbda2f
--- /dev/null
+++ b/apps/fastvideo_studio/src/lib/h3Tags.test.ts
@@ -0,0 +1,117 @@
+import { describe, expect, it } from "vitest";
+import {
+	applyH3Preset,
+	applyH3Tag,
+	H3_PRESETS,
+	H3_TAG_CATEGORIES,
+	H3_TAGS,
+	type H3Preset,
+	type H3Tag,
+} from "@/lib/h3Tags";
+
+const pointTag = H3_TAGS.find((t) => t.id === "pause") as H3Tag;
+const wrapTag = H3_TAGS.find((t) => t.id === "emphasis") as H3Tag;
+const emptyWrapTag = H3_TAGS.find((t) => t.id === "smacks-lips") as H3Tag;
+
+describe("H3 tag catalogue", () => {
+	it("gives every tag a unique id", () => {
+		const ids = H3_TAGS.map((t) => t.id);
+		expect(new Set(ids).size).toBe(ids.length);
+	});
+
+	it("gives wrap tags a close tag and point tags none", () => {
+		for (const tag of H3_TAGS) {
+			if (tag.kind === "wrap") expect(tag.close).toBeTruthy();
+			else expect(tag.close).toBeUndefined();
+		}
+	});
+
+	it("flattens to the same tags the categories declare", () => {
+		const flattened = H3_TAG_CATEGORIES.flatMap((c) => c.tags);
+		expect(H3_TAGS).toEqual(flattened);
+	});
+});
+
+describe("applyH3Tag: point tags", () => {
+	it("inserts at the cursor with padding on both sides", () => {
+		const result = applyH3Tag("Okay, so. This is me.", 9, 9, pointTag);
+		expect(result.value).toBe("Okay, so. <pause> This is me.");
+		// cursor collapses to just past the inserted token
+		expect(result.value.slice(result.start, result.end)).toBe("");
+		expect(result.value.slice(0, result.start)).toBe("Okay, so. <pause> ".slice(0, result.start));
+	});
+
+	it("does not double up whitespace already present", () => {
+		const result = applyH3Tag("Okay, so.  This is me.", 10, 10, pointTag);
+		expect(result.value).toBe("Okay, so. <pause> This is me.");
+	});
+
+	it("inserts cleanly at the very start and end of the text", () => {
+		expect(applyH3Tag("Hello.", 0, 0, pointTag).value).toBe("<pause> Hello.");
+		expect(applyH3Tag("Hello.", 6, 6, pointTag).value).toBe("Hello. <pause>");
+	});
+
+	it("inserts before a selection, keeping the selected text", () => {
+		const result = applyH3Tag("Hello there.", 6, 11, pointTag);
+		expect(result.value).toBe("Hello <pause> there.");
+		expect(result.value.slice(result.start, result.end)).toBe("there");
+	});
+});
+
+describe("applyH3Tag: wrap tags", () => {
+	it("wraps a selection and collapses the cursor after the close tag", () => {
+		const result = applyH3Tag("I was not expecting that.", 6, 9, wrapTag);
+		expect(result.value).toBe("I was <i>not</i> expecting that.");
+		expect(result.start).toBe(result.end);
+		expect(result.value.slice(0, result.start)).toBe("I was <i>not</i>");
+	});
+
+	it("inserts an empty open+close pair with the cursor between them when nothing is selected", () => {
+		const result = applyH3Tag("Don't tell anyone.", 0, 0, wrapTag);
+		expect(result.value).toBe("<i></i>Don't tell anyone.");
+		expect(result.start).toBe("<i>".length);
+		expect(result.end).toBe(result.start);
+	});
+
+	it("supports an intentionally empty pair, as used for a sentence-final effect", () => {
+		const result = applyH3Tag("Okay.", 0, 0, emptyWrapTag);
+		expect(result.value).toBe("<smacks lips></smacks lips>Okay.");
+	});
+});
+
+describe("applyH3Tag: out-of-range selection", () => {
+	it("clamps a selection beyond the text length", () => {
+		const result = applyH3Tag("Hi.", 10, 20, pointTag);
+		expect(result.value).toBe("Hi. <pause>");
+	});
+});
+
+describe("applyH3Preset", () => {
+	const insertPreset = H3_PRESETS.find((p) => p.label === "Nervous") as H3Preset;
+	const wrapPreset = H3_PRESETS.find((p) => p.label === "Intimate") as H3Preset;
+
+	it("inserts with a leading space when the cursor follows text", () => {
+		const r = applyH3Preset("Well", 4, 4, insertPreset);
+		expect(r.value).toBe("Well <uh> <stutter> ");
+		expect(r.start).toBe(r.value.length);
+		expect(r.end).toBe(r.value.length);
+	});
+
+	it("adds no leading space at the start or after whitespace", () => {
+		expect(applyH3Preset("", 0, 0, insertPreset).value).toBe("<uh> <stutter> ");
+		expect(applyH3Preset("Well ", 5, 5, insertPreset).value).toBe("Well <uh> <stutter> ");
+	});
+
+	it("wraps a selection and keeps it selected", () => {
+		const r = applyH3Preset("say it softly now", 4, 13, wrapPreset);
+		expect(r.value).toBe("say <whisper>it softly</whisper> now");
+		expect(r.value.slice(r.start, r.end)).toBe("it softly");
+	});
+
+	it("wraps an empty selection with the caret between the tags", () => {
+		const r = applyH3Preset("hi ", 3, 3, wrapPreset);
+		expect(r.value).toBe("hi <whisper></whisper>");
+		expect(r.start).toBe(r.end);
+		expect(r.value.slice(0, r.start)).toBe("hi <whisper>");
+	});
+});
diff --git a/apps/fastvideo_studio/src/lib/h3Tags.ts b/apps/fastvideo_studio/src/lib/h3Tags.ts
new file mode 100644
index 000000000..d8b414f85
--- /dev/null
+++ b/apps/fastvideo_studio/src/lib/h3Tags.ts
@@ -0,0 +1,409 @@
+/**
+ * MiniMax-H3 speech tags: the inline markers the model recognizes inside
+ * `<d>[English] dialogue</d>` blocks (and elsewhere in a prompt) to shape
+ * delivery -- pauses, breath, emphasis, and non-verbal sounds. `applyH3Tag`
+ * is the pure insertion logic; the tag/category tables drive the console UI.
+ */
+
+export type H3TagKind = "point" | "wrap";
+
+export interface H3Tag {
+	/** stable key for React lists and lookups */
+	id: string;
+	/** short button label, usually the bare tag name */
+	label: string;
+	description: string;
+	/** short usage example shown as a tooltip */
+	example: string;
+	kind: H3TagKind;
+	/** point: the literal token inserted, e.g. "<pause>". wrap: the opening tag. */
+	open: string;
+	/** wrap only: the closing tag, e.g. "</i>". */
+	close?: string;
+	/**
+	 * Rough seconds this tag adds to spoken duration, used only for the
+	 * dialogue editor's length estimate -- ballpark figures, not measured.
+	 * Delivery-only tags (emphasis, whisper, softer) add ~0.
+	 */
+	seconds?: number;
+}
+
+export interface H3TagCategory {
+	id: string;
+	label: string;
+	tags: H3Tag[];
+}
+
+export const H3_TAG_CATEGORIES: H3TagCategory[] = [
+	{
+		id: "pace",
+		label: "Pace & breath",
+		tags: [
+			{
+				id: "pause",
+				label: "<pause>",
+				description: "Short pause",
+				example: "Okay, so. <pause> This is just me talking.",
+				kind: "point",
+				open: "<pause>",
+				seconds: 0.45,
+			},
+			{
+				id: "long-pause",
+				label: "<long pause>",
+				description: "Longer pause",
+				example: "I mean... <long pause> I don't even know.",
+				kind: "point",
+				open: "<long pause>",
+				seconds: 1.2,
+			},
+			{
+				id: "breath",
+				label: "<breath>",
+				description: "Breathing sound",
+				example: "And then... <breath> it just happened.",
+				kind: "point",
+				open: "<breath>",
+				seconds: 0.55,
+			},
+			{
+				id: "inhale",
+				label: "<inhale>",
+				description: "In-breath",
+				example: "<inhale> Alright, let's do this.",
+				kind: "point",
+				open: "<inhale>",
+				seconds: 0.5,
+			},
+			{
+				id: "exhale",
+				label: "<exhale>",
+				description: "Out-breath",
+				example: "<exhale> Okay then.",
+				kind: "point",
+				open: "<exhale>",
+				seconds: 0.6,
+			},
+			{
+				id: "catches-breath",
+				label: "<catches breath>",
+				description: "Out of breath",
+				example: "Wait... <catches breath> hold on a sec.",
+				kind: "point",
+				open: "<catches breath>",
+				seconds: 0.8,
+			},
+			{
+				id: "deep-breath",
+				label: "<deep breath>",
+				description: "Calming down",
+				example: "<deep breath> Okay. I can do this.",
+				kind: "point",
+				open: "<deep breath>",
+				seconds: 1.2,
+			},
+			{
+				id: "pant",
+				label: "<pant>",
+				description: "Panting (single)",
+				example: "Run... <pant> run now!",
+				kind: "point",
+				open: "<pant>",
+				seconds: 0.7,
+			},
+			{
+				id: "pants",
+				label: "<pants>",
+				description: "Panting (continued)",
+				example: "Run... <pants> run now!",
+				kind: "point",
+				open: "<pants>",
+				seconds: 0.7,
+			},
+		],
+	},
+	{
+		id: "delivery",
+		label: "Emphasis & delivery",
+		tags: [
+			{
+				id: "emphasis",
+				label: "<i>…</i>",
+				description: "Emphasize the wrapped words (1-4 words)",
+				example: "I was <i>not</i> expecting that.",
+				kind: "wrap",
+				open: "<i>",
+				close: "</i>",
+				seconds: 0,
+			},
+			{
+				id: "whisper",
+				label: "<whisper>…</whisper>",
+				description: "Whisper delivery",
+				example: "<whisper>Don't tell anyone this.</whisper>",
+				kind: "wrap",
+				open: "<whisper>",
+				close: "</whisper>",
+				seconds: 0,
+			},
+			{
+				id: "softer",
+				label: "<softer>",
+				description: "Quieter delivery",
+				example: "<softer> I don't think I can say it.",
+				kind: "point",
+				open: "<softer>",
+				seconds: 0,
+			},
+			{
+				id: "humming",
+				label: "<humming>…</humming>",
+				description: "Humming a tune",
+				example: "<humming>da-da-da-beautiful-day.</humming>",
+				kind: "wrap",
+				open: "<humming>",
+				close: "</humming>",
+				seconds: 2,
+			},
+		],
+	},
+	{
+		id: "vocal",
+		label: "Vocalizations",
+		tags: [
+			{
+				id: "laughs",
+				label: "<laughs>",
+				description: "Laughing",
+				example: "That's... <laughs> that's actually funny.",
+				kind: "point",
+				open: "<laughs>",
+				seconds: 1.1,
+			},
+			{
+				id: "chuckle",
+				label: "<chuckle>",
+				description: "Laughing (light)",
+				example: "That's... <chuckle> that's actually funny.",
+				kind: "point",
+				open: "<chuckle>",
+				seconds: 0.7,
+			},
+			{
+				id: "sighs",
+				label: "<sighs>",
+				description: "Sigh",
+				example: "<sighs> I really tried.",
+				kind: "point",
+				open: "<sighs>",
+				seconds: 0.85,
+			},
+			{
+				id: "gasp",
+				label: "<gasp>",
+				description: "Sharp intake",
+				example: "<gasp> Oh my God.",
+				kind: "point",
+				open: "<gasp>",
+				seconds: 0.5,
+			},
+			{
+				id: "phew",
+				label: "<phew>",
+				description: "Relief",
+				example: "<phew> That was close.",
+				kind: "point",
+				open: "<phew>",
+				seconds: 0.6,
+			},
+			{
+				id: "mhm",
+				label: "<mhm>",
+				description: "Agreement sound",
+				example: "Yeah, <mhm> exactly.",
+				kind: "point",
+				open: "<mhm>",
+				seconds: 0.4,
+			},
+			{
+				id: "uh",
+				label: "<uh>",
+				description: "Filler / hesitation",
+				example: "So, like... <uh> what was I saying?",
+				kind: "point",
+				open: "<uh>",
+				seconds: 0.35,
+			},
+		],
+	},
+	{
+		id: "body",
+		label: "Disfluency & body",
+		tags: [
+			{
+				id: "stutter",
+				label: "<stutter>",
+				description: "Stutters the words that follow",
+				example: "<stutter> I ca can't believe that.",
+				kind: "point",
+				open: "<stutter>",
+				seconds: 0.45,
+			},
+			{
+				id: "coughs",
+				label: "<coughs>",
+				description: "Cough",
+				example: "<coughs> Sorry, one sec.",
+				kind: "point",
+				open: "<coughs>",
+				seconds: 0.6,
+			},
+			{
+				id: "clears-throat",
+				label: "<clears throat>",
+				description: "Throat clear",
+				example: "<clears throat> So anyway...",
+				kind: "point",
+				open: "<clears throat>",
+				seconds: 0.8,
+			},
+			{
+				id: "sniff",
+				label: "<sniff>",
+				description: "Sniffing",
+				example: "<sniff> It's just... really sad.",
+				kind: "point",
+				open: "<sniff>",
+				seconds: 0.45,
+			},
+			{
+				id: "smacks-lips",
+				label: "<smacks lips>",
+				description:
+					"Lip smack; leave empty to trigger at the end of the sentence, or wrap words to place it there",
+				example: "<smacks lips></smacks lips> Okay.",
+				kind: "wrap",
+				open: "<smacks lips>",
+				close: "</smacks lips>",
+				seconds: 0.3,
+			},
+		],
+	},
+];
+
+export const H3_TAGS: H3Tag[] = H3_TAG_CATEGORIES.flatMap((c) => c.tags);
+
+export interface H3TagApplyResult {
+	value: string;
+	/** selection to restore focus to after the insertion (collapsed unless text was wrapped). */
+	start: number;
+	end: number;
+}
+
+function needsLeadingSpace(before: string): boolean {
+	return before.length > 0 && !/\s$/.test(before);
+}
+
+function needsTrailingSpace(after: string): boolean {
+	return after.length > 0 && !/^\s/.test(after);
+}
+
+/**
+ * Insert `tag` into `value` at the given selection.
+ *
+ * Point tags are inserted at the selection start, padded with a space on
+ * either side where the surrounding text doesn't already have one, and the
+ * cursor collapses to just past the inserted token.
+ *
+ * Wrap tags wrap the current selection (cursor lands just past the closing
+ * tag), or insert an empty `open+close` pair with the cursor left between
+ * them when nothing is selected, ready to type into.
+ */
+export function applyH3Tag(
+	value: string,
+	selectionStart: number,
+	selectionEnd: number,
+	tag: H3Tag,
+): H3TagApplyResult {
+	const start = Math.max(0, Math.min(selectionStart, value.length));
+	const end = Math.max(start, Math.min(selectionEnd, value.length));
+	const before = value.slice(0, start);
+	const selected = value.slice(start, end);
+	const after = value.slice(end);
+
+	if (tag.kind === "point") {
+		const lead = needsLeadingSpace(before) ? " " : "";
+		const trailingContext = selected.length > 0 ? selected : after;
+		const trail = needsTrailingSpace(trailingContext) ? " " : "";
+		const token = `${lead}${tag.open}${trail}`;
+		const cursor = before.length + token.length;
+		return {
+			value: before + token + selected + after,
+			start: cursor,
+			end: cursor + selected.length,
+		};
+	}
+
+	const close = tag.close ?? "";
+	if (selected.length > 0) {
+		const cursor = before.length + tag.open.length + selected.length + close.length;
+		return {
+			value: before + tag.open + selected + close + after,
+			start: cursor,
+			end: cursor,
+		};
+	}
+	const cursor = before.length + tag.open.length;
+	return {
+		value: before + tag.open + close + after,
+		start: cursor,
+		end: cursor,
+	};
+}
+
+/** One-click bundles of tags for common delivery beats (dialogue editor + shot list). */
+export interface H3Preset {
+	label: string;
+	/** literal text inserted at the cursor (a leading space is added when needed). */
+	insert?: string;
+	/** wrap the selection between these two tokens instead of inserting. */
+	wrap?: [string, string];
+}
+
+export const H3_PRESETS: H3Preset[] = [
+	{ label: "Cracking", insert: "<catches breath> <stutter> " },
+	{ label: "Nervous", insert: "<uh> <stutter> " },
+	{ label: "Relief", insert: "<phew> <exhale> " },
+	{ label: "Cold turn", insert: "<long pause> <softer> " },
+	{ label: "Intimate", wrap: ["<whisper>", "</whisper>"] },
+	{ label: "Amused", insert: "<chuckle> " },
+];
+
+/** Pure counterpart of `applyH3Tag` for presets: same {value, selection} contract. */
+export function applyH3Preset(
+	value: string,
+	selectionStart: number,
+	selectionEnd: number,
+	preset: H3Preset,
+): H3TagApplyResult {
+	const start = Math.max(0, Math.min(selectionStart, value.length));
+	const end = Math.max(start, Math.min(selectionEnd, value.length));
+	const before = value.slice(0, start);
+	const mid = value.slice(start, end);
+	const after = value.slice(end);
+
+	if (preset.wrap) {
+		const [open, close] = preset.wrap;
+		const innerStart = before.length + open.length;
+		return {
+			value: before + open + mid + close + after,
+			start: innerStart,
+			end: innerStart + mid.length,
+		};
+	}
+
+	const lead = before.length && !/\s$/.test(before) ? " " : "";
+	const chunk = lead + (preset.insert ?? "");
+	const caret = before.length + chunk.length;
+	return { value: before + chunk + after, start: caret, end: caret };
+}
diff --git a/apps/fastvideo_studio/src/lib/jobJson.test.ts b/apps/fastvideo_studio/src/lib/jobJson.test.ts
new file mode 100644
index 000000000..ddefbc10f
--- /dev/null
+++ b/apps/fastvideo_studio/src/lib/jobJson.test.ts
@@ -0,0 +1,78 @@
+import { describe, expect, it } from "vitest";
+import { parseJobJson } from "@/lib/jobJson";
+
+const ok = (text: string) => {
+	const r = parseJobJson(text);
+	if (!r.ok) throw new Error(r.error);
+	return r;
+};
+
+describe("parseJobJson", () => {
+	it("joins a prompt given as lines and carries the settings over", () => {
+		const r = ok(
+			JSON.stringify({
+				model_id: "MiniMaxAI/MiniMax-H3",
+				name: "demo",
+				workload_type: "i2v",
+				prompt: ["summary:", "Hello.", "", "detailed_description:", "[Shot 1] A."],
+				references: [{ source: "/data/a.png", media_type: "image" }],
+				guidance_scale: 1,
+				num_frames: 243,
+				fps: 24,
+				num_gpus: 4,
+				use_fsdp_inference: true,
+			}),
+		);
+		expect(r.workloadType).toBe("i2v");
+		expect(r.job.prompt).toBe("summary:\nHello.\n\ndetailed_description:\n[Shot 1] A.");
+		expect(r.job).toMatchObject({
+			model_id: "MiniMaxAI/MiniMax-H3",
+			name: "demo",
+			guidance_scale: 1,
+			num_frames: 243,
+			num_gpus: 4,
+			use_fsdp_inference: true,
+			references: [{ source: "/data/a.png", media_type: "image" }],
+		});
+	});
+
+	it("accepts a plain string prompt", () => {
+		expect(ok(JSON.stringify({ model_id: "m", prompt: "one line" })).job.prompt).toBe("one line");
+	});
+
+	it("infers the workload when it isn't declared", () => {
+		expect(ok(JSON.stringify({ model_id: "m", prompt: "p" })).workloadType).toBe("t2v");
+		expect(
+			ok(JSON.stringify({ model_id: "m", prompt: "p", references: [{ source: "/a/b.mp4" }] })).workloadType,
+		).toBe("i2v");
+		expect(ok(JSON.stringify({ model_id: "m", prompt: "p", image_path: "/a.png" })).workloadType).toBe("i2v");
+	});
+
+	it("infers a reference's media type from its extension", () => {
+		const r = ok(
+			JSON.stringify({
+				model_id: "m",
+				prompt: "p",
+				references: [{ source: "/a/b.mp4" }, { source: "/a/c.wav" }, { source: "/a/d.png" }],
+			}),
+		);
+		expect(r.job.references?.map((x) => x.media_type)).toEqual(["video", "audio", "image"]);
+	});
+
+	it("ignores settings of the wrong type instead of passing them through", () => {
+		const r = ok(JSON.stringify({ model_id: "m", prompt: "p", num_frames: "many", use_fsdp_inference: "yes", extra: 1 }));
+		expect(r.job.num_frames).toBeUndefined();
+		expect(r.job.use_fsdp_inference).toBeUndefined();
+		expect(r.job).not.toHaveProperty("extra");
+	});
+
+	it.each([
+		["not json", "That file isn't valid JSON."],
+		["[1,2]", "Expected a JSON object describing one job."],
+		[JSON.stringify({ prompt: "p" }), "The job file has no model_id."],
+		[JSON.stringify({ model_id: "m", prompt: 5 }), "prompt must be text or a list of text lines."],
+		[JSON.stringify({ model_id: "m", prompt: "p", references: [{}] }), "references[0] needs a source path."],
+	])("rejects %s", (text, error) => {
+		expect(parseJobJson(text)).toEqual({ ok: false, error });
+	});
+});
diff --git a/apps/fastvideo_studio/src/lib/jobJson.ts b/apps/fastvideo_studio/src/lib/jobJson.ts
new file mode 100644
index 000000000..62a88f3db
--- /dev/null
+++ b/apps/fastvideo_studio/src/lib/jobJson.ts
@@ -0,0 +1,102 @@
+/**
+ * Read a job file (the `ref2va_*.json` shape) into the create-job form's
+ * job shape. Keys match the API's CreateJobRequest; `prompt` may be a string
+ * or a list of lines (joined with newlines), since job files keep it as lines
+ * for readability.
+ */
+import type { JobLike } from "@/lib/jobToFields";
+
+export type JobJsonResult =
+	| { ok: true; job: JobLike; workloadType: "t2v" | "i2v" | "t2i" }
+	| { ok: false; error: string };
+
+const WORKLOADS = ["t2v", "i2v", "t2i"] as const;
+const MEDIA_TYPES = ["image", "video", "audio"] as const;
+const VIDEO_EXT = /\.(mp4|mov|mkv|webm|avi)$/i;
+const AUDIO_EXT = /\.(wav|mp3|flac|m4a|ogg)$/i;
+
+const STRING_KEYS = ["name", "image_path", "last_image_path", "negative_prompt"] as const;
+const NUMBER_KEYS = [
+	"num_inference_steps",
+	"num_frames",
+	"height",
+	"width",
+	"guidance_scale",
+	"guidance_rescale",
+	"fps",
+	"seed",
+	"num_gpus",
+	"vsa_sparsity",
+	"tp_size",
+	"sp_size",
+] as const;
+const BOOLEAN_KEYS = [
+	"dit_cpu_offload",
+	"dit_layerwise_offload",
+	"text_encoder_cpu_offload",
+	"vae_cpu_offload",
+	"image_encoder_cpu_offload",
+	"use_fsdp_inference",
+	"enable_torch_compile",
+] as const;
+
+function inferMediaType(source: string): (typeof MEDIA_TYPES)[number] {
+	if (VIDEO_EXT.test(source)) return "video";
+	if (AUDIO_EXT.test(source)) return "audio";
+	return "image";
+}
+
+export function parseJobJson(text: string): JobJsonResult {
+	let raw: unknown;
+	try {
+		raw = JSON.parse(text);
+	} catch {
+		return { ok: false, error: "That file isn't valid JSON." };
+	}
+	if (typeof raw !== "object" || raw === null || Array.isArray(raw)) {
+		return { ok: false, error: "Expected a JSON object describing one job." };
+	}
+	const data = raw as Record<string, unknown>;
+
+	if (typeof data.model_id !== "string" || !data.model_id.trim()) {
+		return { ok: false, error: "The job file has no model_id." };
+	}
+
+	let prompt: string;
+	if (typeof data.prompt === "string") prompt = data.prompt;
+	else if (Array.isArray(data.prompt) && data.prompt.every((l) => typeof l === "string")) {
+		prompt = (data.prompt as string[]).join("\n");
+	} else {
+		return { ok: false, error: "prompt must be text or a list of text lines." };
+	}
+
+	const references: { source: string; media_type: string }[] = [];
+	if (data.references !== undefined && data.references !== null) {
+		if (!Array.isArray(data.references)) return { ok: false, error: "references must be a list." };
+		for (const [i, ref] of data.references.entries()) {
+			const r = ref as Record<string, unknown> | null;
+			if (!r || typeof r.source !== "string" || !r.source) {
+				return { ok: false, error: `references[${i}] needs a source path.` };
+			}
+			const mediaType =
+				typeof r.media_type === "string" && (MEDIA_TYPES as readonly string[]).includes(r.media_type)
+					? r.media_type
+					: inferMediaType(r.source);
+			references.push({ source: r.source, media_type: mediaType });
+		}
+	}
+
+	const job: JobLike = { id: `imported-${Date.now()}`, model_id: data.model_id.trim(), prompt, references };
+	const target = job as unknown as Record<string, unknown>;
+	for (const key of STRING_KEYS) if (typeof data[key] === "string") target[key] = data[key];
+	for (const key of NUMBER_KEYS) if (typeof data[key] === "number" && Number.isFinite(data[key])) target[key] = data[key];
+	for (const key of BOOLEAN_KEYS) if (typeof data[key] === "boolean") target[key] = data[key];
+
+	const declared = (WORKLOADS as readonly string[]).includes(data.workload_type as string)
+		? (data.workload_type as (typeof WORKLOADS)[number])
+		: null;
+	const workloadType = declared ?? (references.length > 0 || job.image_path ? "i2v" : "t2v");
+	job.workload_type = workloadType;
+
+	return { ok: true, job, workloadType };
+}
diff --git a/apps/fastvideo_studio/src/lib/scenes.test.ts b/apps/fastvideo_studio/src/lib/scenes.test.ts
new file mode 100644
index 000000000..0b1f41559
--- /dev/null
+++ b/apps/fastvideo_studio/src/lib/scenes.test.ts
@@ -0,0 +1,250 @@
+import { describe, expect, it } from "vitest";
+import { UNNAMED_SCENE, buildScenes, clipSeconds, formatClock, parseSceneName, shotsOfJob } from "@/lib/scenes";
+import { shotLines, shotProse } from "@/lib/h3Shots";
+import type { Job } from "@/lib/types";
+
+function job(over: Partial<Job> & { id: string }): Job {
+	return {
+		model_id: "m",
+		prompt: "",
+		status: "pending",
+		created_at: 0,
+		started_at: null,
+		finished_at: null,
+		error: null,
+		output_path: null,
+		log_file_path: null,
+		num_inference_steps: 50,
+		num_frames: 240,
+		fps: 24,
+		height: 768,
+		width: 1344,
+		guidance_scale: 1,
+		seed: 0,
+		num_gpus: 1,
+		progress: 0,
+		progress_msg: "",
+		phase: "",
+		...over,
+	};
+}
+
+const SECTIONS = (detailed: string, subjects = "<Subject 1> is a man.") =>
+	["subject_definitions:", subjects, "", "detailed_description:", detailed].join("\n");
+
+describe("parseSceneName", () => {
+	it.each([
+		["wolf-lunch-clip-01", "wolf-lunch", 1],
+		["wolf-lunch-clip-15", "wolf-lunch", 15],
+		["closeup-fictional-actor-monologue-shot2", "closeup-fictional-actor-monologue", 2],
+		["Rooftop Standoff part 3", "Rooftop Standoff", 3],
+		["scene_a_07", "scene_a", 7],
+	])("%s -> stem %s, clip %i", (name, title, order) => {
+		expect(parseSceneName(name)).toEqual({ key: title.toLowerCase(), title, order });
+	});
+
+	it("treats a name with no trailing number as its own scene", () => {
+		expect(parseSceneName("Rooftop Standoff")).toEqual({ key: "rooftop standoff", title: "Rooftop Standoff", order: null });
+	});
+
+	it("keeps the word scene as part of the name", () => {
+		expect(parseSceneName("rooftop-scene-2")).toEqual({ key: "rooftop-scene", title: "rooftop-scene", order: 2 });
+	});
+
+	it("does not treat a bare number or number-only suffix as a stem", () => {
+		expect(parseSceneName("42").order).toBeNull();
+		expect(parseSceneName("clip-3").order).toBeNull();
+	});
+
+	it("puts blank or missing names in the unnamed group", () => {
+		expect(parseSceneName(undefined).key).toBe(UNNAMED_SCENE);
+		expect(parseSceneName("   ").key).toBe(UNNAMED_SCENE);
+	});
+
+	it("groups names that differ only in case", () => {
+		expect(parseSceneName("Wolf-Lunch-Clip-1").key).toBe(parseSceneName("wolf-lunch-clip-2").key);
+	});
+});
+
+describe("shotsOfJob", () => {
+	it("reads shots and dialogue out of a six-section prompt", () => {
+		const shots = shotsOfJob({
+			prompt: SECTIONS(
+				"[Shot 1] Wide shot. A dock.\n[Shot 2] At 00:02.000, the shot cuts to a close-up shot. (S1) <d>[English] Hi.</d>",
+			),
+		});
+		expect(shots).toHaveLength(2);
+		expect(shots[1].type).toBe("close-up");
+		expect(shotLines(shots[1])[0].text).toBe("Hi.");
+		expect(shots[0].subjectIds).toEqual([]);
+	});
+
+	it("reads shots out of a base-format integrated_multimodal_description", () => {
+		const shots = shotsOfJob({
+			prompt: "For the target video, <Picture 1> is used.\n\nintegrated_multimodal_description: [Shot 1] He (S1) says: <d>[English] Sorry.</d> <cutoff>\n\noverall_soundscape: Quiet.",
+		});
+		expect(shots).toHaveLength(1);
+		expect(shotLines(shots[0]).map((l) => l.text)).toEqual(["Sorry."]);
+	});
+
+	it("shows a plain prompt as a single shot", () => {
+		const shots = shotsOfJob({ prompt: "A toy car drives into a plush dog." });
+		expect(shots).toHaveLength(1);
+		expect(shotProse(shots[0])).toBe("A toy car drives into a plush dog.");
+	});
+
+	it("returns no shots for an empty prompt", () => {
+		expect(shotsOfJob({ prompt: "" })).toEqual([]);
+	});
+});
+
+describe("buildScenes", () => {
+	const three = [
+		job({ id: "c", name: "wolf-lunch-clip-03", created_at: 3, status: "pending", prompt: "[Shot 1] Wide shot. C." }),
+		job({ id: "a", name: "wolf-lunch-clip-01", created_at: 1, status: "completed", num_frames: 240, prompt: "[Shot 1] Wide shot. A.\n[Shot 2] At 00:02.000, the shot cuts to a close-up shot. A2." }),
+		job({ id: "b", name: "wolf-lunch-clip-02", created_at: 2, status: "completed", num_frames: 120, fps: 24, prompt: "[Shot 1] Close-up shot. B." }),
+	];
+
+	it("groups by stem and orders clips by their number, not creation order", () => {
+		const [scene] = buildScenes(three);
+		expect(scene.title).toBe("wolf-lunch");
+		expect(scene.clips.map((c) => c.job.id)).toEqual(["a", "b", "c"]);
+		expect(scene.clips.map((c) => c.position)).toEqual([1, 2, 3]);
+	});
+
+	it("lays every shot on one timeline using each clip's frames and fps", () => {
+		const [scene] = buildScenes(three);
+		expect(scene.clips.map((c) => c.startSeconds)).toEqual([0, 10, 15]);
+		expect(scene.totalSeconds).toBe(25); // 240/24 + 120/24 + 240/24 = 10 + 5 + 10
+	});
+
+	it("counts shots and statuses across the scene", () => {
+		const [scene] = buildScenes(three);
+		expect(scene.shotCount).toBe(4);
+		expect(scene.statusCounts).toEqual({ completed: 2, pending: 1 });
+	});
+
+	it("offsets each shot's cut time by the clip's start", () => {
+		const [scene] = buildScenes(three);
+		const second = scene.clips[1];
+		expect(second.shotStarts).toEqual([10]);
+		const first = scene.clips[0];
+		expect(first.shotStarts[0]).toBe(0);
+		expect(first.shotStarts[1]).toBe(2);
+	});
+
+	it("keeps separate scenes apart, newest first, with unnamed jobs last", () => {
+		const scenes = buildScenes([
+			job({ id: "1", name: "old-scene-1", created_at: 5 }),
+			job({ id: "2", name: "new-scene-1", created_at: 50 }),
+			job({ id: "3", name: "", created_at: 100 }),
+		]);
+		expect(scenes.map((s) => s.title)).toEqual(["new-scene", "old-scene", "Unnamed jobs"]);
+	});
+
+	it("puts clips without a number after numbered ones, oldest first", () => {
+		const [scene] = buildScenes([
+			job({ id: "x", name: "wolf-1", created_at: 9 }),
+			job({ id: "y", name: "wolf", created_at: 1 }),
+			job({ id: "z", name: "wolf-2", created_at: 2 }),
+		]);
+		// "wolf" shares the stem with wolf-1 / wolf-2, so it joins them -- last, having no number
+		expect(scene.clips.map((c) => c.job.id)).toEqual(["x", "z", "y"]);
+	});
+
+	it("flags a clip that continues the previous shot", () => {
+		const [scene] = buildScenes([
+			job({ id: "a", name: "s-1", prompt: "[Shot 1] Close-up shot. A speech. <cutoff>" }),
+			job({ id: "b", name: "s-2", prompt: "[Shot 1] Close-up shot, dolly. Continuing the same shot on <Subject 1>, still speaking." }),
+			job({ id: "c", name: "s-3", prompt: "summary:\n[reference generation] It continues the previous shot without a cut.\n\ndetailed_description:\n[Shot 1] Wide shot. X." }),
+			job({ id: "d", name: "s-4", prompt: "[Shot 1] Wide shot. Fresh." }),
+		]);
+		expect(scene.clips.map((c) => c.continuation)).toEqual([false, true, true, false]);
+	});
+
+	it("falls back to 24 fps when a job has none", () => {
+		expect(clipSeconds({ num_frames: 240, fps: undefined })).toBe(10);
+		expect(clipSeconds({ num_frames: 240, fps: 0 })).toBe(10);
+		expect(clipSeconds({ num_frames: 240, fps: 30 })).toBe(8);
+	});
+
+	it("returns no scenes for no jobs", () => {
+		expect(buildScenes([])).toEqual([]);
+	});
+});
+
+describe("formatClock", () => {
+	it("formats m:ss and h:mm:ss", () => {
+		expect(formatClock(0)).toBe("0:00");
+		expect(formatClock(9.6)).toBe("0:10");
+		expect(formatClock(65)).toBe("1:05");
+		expect(formatClock(3725)).toBe("1:02:05");
+		expect(formatClock(-4)).toBe("0:00");
+	});
+});
+
+describe("takes of the same clip", () => {
+	const take = (id: string, n: number, over: Partial<Job> = {}) =>
+		job({ id, name: `wolf-lunch-clip-0${n}`, prompt: `[Shot 1] Wide shot. ${id}.`, num_frames: 240, ...over });
+
+	it("counts re-runs of a clip number once, not as extra clips", () => {
+		const [scene] = buildScenes([
+			take("a1", 1, { created_at: 1, status: "completed" }),
+			take("a2", 1, { created_at: 2, status: "completed" }),
+			take("a3", 1, { created_at: 3, status: "completed" }),
+			take("b1", 2, { created_at: 4, status: "completed" }),
+		]);
+		expect(scene.clips).toHaveLength(2);
+		expect(scene.totalSeconds).toBe(20);
+		expect(scene.shotCount).toBe(2);
+		expect(scene.statusCounts).toEqual({ completed: 2 });
+		expect(scene.clips[0].takes.map((j) => j.id)).toEqual(["a1", "a2", "a3"]);
+	});
+
+	it("shows the most recent completed take, even over a newer failed one", () => {
+		const [scene] = buildScenes([
+			take("old", 1, { created_at: 1, status: "completed", finished_at: 10 }),
+			take("good", 1, { created_at: 2, status: "completed", finished_at: 20 }),
+			take("bad", 1, { created_at: 3, status: "failed", finished_at: 30 }),
+		]);
+		expect(scene.clips[0].job.id).toBe("good");
+	});
+
+	it("falls back to the newest take when none completed", () => {
+		const [scene] = buildScenes([
+			take("p1", 1, { created_at: 1, status: "failed" }),
+			take("p2", 1, { created_at: 2, status: "pending" }),
+		]);
+		expect(scene.clips[0].job.id).toBe("p2");
+	});
+
+	it("shows the take the user chose, and lays the timeline out from it", () => {
+		const jobs = [
+			take("short", 1, { created_at: 1, status: "completed", num_frames: 120 }),
+			take("long", 1, { created_at: 2, status: "completed", num_frames: 480 }),
+			take("next", 2, { created_at: 3 }),
+		];
+		const [auto] = buildScenes(jobs);
+		expect(auto.clips[0].job.id).toBe("long");
+		expect(auto.clips[1].startSeconds).toBe(20);
+
+		const [chosen] = buildScenes(jobs, { [auto.clips[0].takeKey as string]: "short" });
+		expect(chosen.clips[0].job.id).toBe("short");
+		expect(chosen.clips[1].startSeconds).toBe(5);
+	});
+
+	it("ignores a choice that isn't one of the takes", () => {
+		const jobs = [take("a", 1, { status: "completed" })];
+		const [scene] = buildScenes(jobs, { "wolf-lunch#1": "no-such-job" });
+		expect(scene.clips[0].job.id).toBe("a");
+	});
+
+	it("keeps each unnumbered job as its own clip", () => {
+		const [scene] = buildScenes([
+			job({ id: "u1", name: "solo", created_at: 1 }),
+			job({ id: "u2", name: "solo", created_at: 2 }),
+		]);
+		expect(scene.clips.map((c) => c.job.id)).toEqual(["u1", "u2"]);
+		expect(scene.clips.every((c) => c.takeKey === null)).toBe(true);
+	});
+});
diff --git a/apps/fastvideo_studio/src/lib/scenes.ts b/apps/fastvideo_studio/src/lib/scenes.ts
new file mode 100644
index 000000000..2092003b7
--- /dev/null
+++ b/apps/fastvideo_studio/src/lib/scenes.ts
@@ -0,0 +1,182 @@
+/**
+ * Group jobs into scenes and lay their shots out on one timeline.
+ *
+ * A long scene is generated as many short clips, one job each. Nothing on the
+ * server links them, so the link is the job name: `wolf-lunch-clip-01`,
+ * `wolf-lunch-clip-02`, ... share the stem `wolf-lunch` and are ordered by the
+ * trailing number. Everything else -- shots, dialogue, subjects -- is read back
+ * out of each job's own prompt, so no backend support is needed.
+ */
+import { readField, shotTarget } from "@/lib/h3PromptFields";
+import {
+	parseDetailedDescription,
+	parseSubjectLabels,
+	resolveShotStarts,
+	shotProse,
+	type Shot,
+} from "@/lib/h3Shots";
+import type { Job } from "@/lib/types";
+
+export const UNNAMED_SCENE = "__unnamed__";
+const DEFAULT_FPS = 24;
+
+// "<stem>[-_. ][clip|part|pt|shot|take][-_. ]<number>". "scene" is deliberately not
+// a clip word: it is far more likely part of the scene's own name ("rooftop-scene-2").
+const CLIP_SUFFIX_RE = /^(.*?)[\s._-]*(?:clip|part|pt|shot|take)?[\s._-]*(\d+)$/i;
+
+export interface SceneName {
+	key: string;
+	title: string;
+	/** the clip number from the end of the name, or null when there isn't one */
+	order: number | null;
+}
+
+export function parseSceneName(name: string | undefined | null): SceneName {
+	const trimmed = (name ?? "").trim();
+	if (!trimmed) return { key: UNNAMED_SCENE, title: "Unnamed jobs", order: null };
+	const m = trimmed.match(CLIP_SUFFIX_RE);
+	const stem = m ? m[1].replace(/[\s._-]+$/, "") : "";
+	if (m && stem) return { key: stem.toLowerCase(), title: stem, order: Number(m[2]) };
+	return { key: trimmed.toLowerCase(), title: trimmed, order: null };
+}
+
+/** The shots a job's prompt describes, wherever they live in it (base, reference, or plain prompt). */
+export function shotsOfJob(job: Pick<Job, "prompt">): Shot[] {
+	const prompt = job.prompt ?? "";
+	const labels = parseSubjectLabels(readField(prompt, "subject_definitions") ?? "");
+	return parseDetailedDescription(shotTarget(prompt).body, labels).shots;
+}
+
+export interface SceneClip {
+	/** the take shown for this clip */
+	job: Job;
+	/**
+	 * Every job that claims this clip number, oldest first. Re-running a clip
+	 * gives the same name again, so these are takes of one clip, not extra clips.
+	 */
+	takes: Job[];
+	/** identifies this clip for choosing a take; null for a clip with no number */
+	takeKey: string | null;
+	/** 1-based place in the scene */
+	position: number;
+	shots: Shot[];
+	/** seconds from the start of the scene to the start of this clip */
+	startSeconds: number;
+	durationSeconds: number;
+	/** each shot's cut-in time measured from the start of the scene */
+	shotStarts: number[];
+	/** this clip picks up in the middle of the previous clip's last shot */
+	continuation: boolean;
+}
+
+export interface Scene {
+	key: string;
+	title: string;
+	clips: SceneClip[];
+	totalSeconds: number;
+	shotCount: number;
+	statusCounts: Record<string, number>;
+	/** most recent created_at among the clips, for ordering scenes */
+	latest: number;
+}
+
+export function clipSeconds(job: Pick<Job, "num_frames" | "fps">): number {
+	const fps = job.fps && job.fps > 0 ? job.fps : DEFAULT_FPS;
+	return Math.max(0, (job.num_frames ?? 0) / fps);
+}
+
+function isContinuation(job: Job, shots: Shot[]): boolean {
+	if (/continues the previous shot/i.test(job.prompt ?? "")) return true;
+	return shots.length > 0 && /^Continuing the same shot/i.test(shotProse(shots[0]));
+}
+
+function createdAt(job: Job): number {
+	return job.created_at ?? 0;
+}
+
+/** The take to show: the one the user chose, else the most recent completed one, else the newest. */
+function pickTake(takes: Job[], choice?: string): Job {
+	const chosen = choice ? takes.find((j) => j.id === choice) : undefined;
+	if (chosen) return chosen;
+	const recent = (a: Job, b: Job) => (b.finished_at ?? createdAt(b)) - (a.finished_at ?? createdAt(a));
+	const completed = takes.filter((j) => j.status === "completed").sort(recent);
+	return completed[0] ?? [...takes].sort((a, b) => createdAt(b) - createdAt(a))[0];
+}
+
+/**
+ * Group inference jobs into scenes, each with its clips in order and a
+ * scene-wide timeline. `choices` maps a clip's `takeKey` to the job id of the
+ * take to show in place of the default.
+ */
+export function buildScenes(jobs: Job[], choices: Record<string, string> = {}): Scene[] {
+	const groups = new Map<string, { title: string; entries: { job: Job; order: number | null }[] }>();
+	for (const job of jobs) {
+		const { key, title, order } = parseSceneName(job.name);
+		const group = groups.get(key) ?? { title, entries: [] };
+		group.entries.push({ job, order });
+		groups.set(key, group);
+	}
+
+	const scenes: Scene[] = [];
+	for (const [key, { title, entries }] of groups) {
+		// One clip per clip number (its takes together); each unnumbered job stands alone, after the numbered ones.
+		const byNumber = new Map<number, Job[]>();
+		const unnumbered: Job[][] = [];
+		for (const { job, order } of entries) {
+			if (order === null) unnumbered.push([job]);
+			else byNumber.set(order, [...(byNumber.get(order) ?? []), job]);
+		}
+		const slots: { takes: Job[]; takeKey: string | null }[] = [
+			...[...byNumber.entries()].sort(([a], [b]) => a - b).map(([order, takes]) => ({ takes, takeKey: `${key}#${order}` })),
+			...unnumbered.sort((a, b) => createdAt(a[0]) - createdAt(b[0])).map((takes) => ({ takes, takeKey: null })),
+		];
+
+		let elapsed = 0;
+		const statusCounts: Record<string, number> = {};
+		const clips: SceneClip[] = slots.map(({ takes, takeKey }, i) => {
+			const ordered = [...takes].sort((a, b) => createdAt(a) - createdAt(b));
+			const job = pickTake(ordered, takeKey ? choices[takeKey] : undefined);
+			const shots = shotsOfJob(job);
+			const durationSeconds = clipSeconds(job);
+			const clip: SceneClip = {
+				job,
+				takes: ordered,
+				takeKey,
+				position: i + 1,
+				shots,
+				startSeconds: elapsed,
+				durationSeconds,
+				shotStarts: resolveShotStarts(shots).map((t) => elapsed + t),
+				continuation: isContinuation(job, shots),
+			};
+			elapsed += durationSeconds;
+			statusCounts[job.status] = (statusCounts[job.status] ?? 0) + 1;
+			return clip;
+		});
+
+		scenes.push({
+			key,
+			title,
+			clips,
+			totalSeconds: elapsed,
+			shotCount: clips.reduce((n, c) => n + c.shots.length, 0),
+			statusCounts,
+			latest: Math.max(0, ...entries.map((e) => createdAt(e.job))),
+		});
+	}
+
+	// Newest scene first; the catch-all group of unnamed jobs goes last.
+	return scenes.sort((a, b) => {
+		if ((a.key === UNNAMED_SCENE) !== (b.key === UNNAMED_SCENE)) return a.key === UNNAMED_SCENE ? 1 : -1;
+		return b.latest - a.latest;
+	});
+}
+
+/** m:ss, or h:mm:ss past an hour -- for clip ranges and runtimes. */
+export function formatClock(seconds: number): string {
+	const total = Math.max(0, Math.round(seconds));
+	const h = Math.floor(total / 3600);
+	const m = Math.floor((total % 3600) / 60);
+	const sec = String(total % 60).padStart(2, "0");
+	return h > 0 ? `${h}:${String(m).padStart(2, "0")}:${sec}` : `${m}:${sec}`;
+}
diff --git a/apps/fastvideo_studio/src/lib/types.ts b/apps/fastvideo_studio/src/lib/types.ts
index a0d99ae1b..b7bc907f9 100644
--- a/apps/fastvideo_studio/src/lib/types.ts
+++ b/apps/fastvideo_studio/src/lib/types.ts
@@ -18,6 +18,8 @@ export interface Job {
 	log_file_path: string | null;
 	num_inference_steps: number;
 	num_frames: number;
+	/** frames per second; the API sends it, older cached objects may not have it */
+	fps?: number;
 	height: number;
 	width: number;
 	guidance_scale: number;
diff --git a/apps/fastvideo_studio/src/test/setup.ts b/apps/fastvideo_studio/src/test/setup.ts
index 35dc4b6ea..1886e8f8f 100644
--- a/apps/fastvideo_studio/src/test/setup.ts
+++ b/apps/fastvideo_studio/src/test/setup.ts
@@ -16,6 +16,23 @@ if (!globalThis.crypto?.randomUUID) {
   });
 }
 
+// jsdom's Blob/File predate Blob.text(); real browsers all have it.
+if (!Blob.prototype.text) {
+  Blob.prototype.text = function text(this: Blob) {
+    return new Promise<string>((resolve, reject) => {
+      const reader = new FileReader();
+      reader.onload = () => resolve(String(reader.result));
+      reader.onerror = () => reject(reader.error);
+      reader.readAsText(this);
+    });
+  };
+}
+
+// jsdom doesn't implement scrollIntoView.
+if (!Element.prototype.scrollIntoView) {
+  Element.prototype.scrollIntoView = vi.fn();
+}
+
 if (!URL.createObjectURL) {
   URL.createObjectURL = vi.fn(() => 'blob:mock-url');
 }